# Harness capability map — full validation checklist As of 2026-09-15. Scoring: **2** = first-party documented page/command · **1** = plugin / MCP / community / preview only · **0** = not found in that project's own docs. A "first-party page" means *any* vendor-published page — product/download/install pages count, not only `/docs/*` paths. Max 54 per harness = 27 capabilities × 2. Every cell below needs a verdict, not just the ones in dispute. Reply as: `harness · capability · we said X · should be Y · source URL` (or `confirmed`). **v4.2 revision (compiled 2026-09-16 — round 2 closed: 2 corrections accepted, 1 rejected, 1 citation repaired):** - OpenClaw 2.0 · *Exposes itself over MCP* **0 → 2** (@Gemini; https://docs.openclaw.ai/cli/mcp/serve — a dedicated page for `openclaw mcp serve`, a stdio MCP server exposing routed conversations, transcripts, `events_poll`/`events_wait` and approval tools, with a Docker smoke test. This was the cell we flagged as "if a serve surface exists, this is the cell that moves". It exists.) - OpenClaw 1.x · *Generates images / audio / video* **1 → 2** (@Gemini and @GLM both dated video/music generation to a pre-2.0 release — https://github.com/openclaw/openclaw/releases/tag/v2026.4.5 ships the built-in `video_generate` and `music_generate` tools on 2026-04-06, and the v2026.7.1 notes then harden Veo/OpenAI/BytePlus/OpenRouter video paths. Our published condition was "date it and it goes to 2". @GLM's own evidence met that condition while their verdict said "stays at 1" — the rule decides.) - OpenClaw 1.x · *Multi-agent orchestration* — **correction rejected.** The claim was that Swarm, Fleets and Task Flow all arrived in 2.0, with "0 matches" in v2026.7.1. Swarm and Fleets do check out as 2.0 (0 mentions in the 1.x notes; their PRs sit above the 1.x-era ceiling), but Task Flow does not: the v2026.7.1 notes reference TaskFlow validation, recoverable TaskFlow state and linked TaskFlows three times, *Workboard* appears 13 times, and the current Task Flow page calls itself "formerly ClawFlow" — a 1.x-era name. A Control UI board plus a flow layer is a 2 under our own rubric (we scored Hermes 2 on a kanban board), so 1.x stays at 2. What 2.0 added is Swarm collector children and Fleets *on top of* an orchestration layer that already existed. - Citation repair, no score change: OpenClaw 2.0 *Browser & screen control* was sourced to `/tools/screen`, which documents Control UI pane layout and navigation and says outright that it is "not screenshot capture or browser automation". Repinned to https://docs.openclaw.ai/tools/browser. - **Net:** OpenClaw 2.0 **49/54** · OpenClaw 1.x **47/54** — the 16,000-PR delta is **two points**, not one, and it is two points for corrected reasons (one under-credited cell in each column). The "rebuild, not expansion" explanation does not survive: Swarm/Fleets is a new coordination architecture and the bundled IMAP watcher is a new inbound-trigger surface, both absent from the 1.x notes. Two of the four categories claimed as new do not survive checking — see `corrections.md`. **v3.2 revision (compiled 2026-09-15 — Pi added as a 10th column; no existing cell moved):** - **Pi (pi.dev) joins at 26/54**, scored from its own docs tree (`packages/coding-agent/docs/*` + the first-party package catalog), not from the marketing page. `@earendil-works/pi-coding-agent`, MIT, by Mario Zechner / Earendil Inc. The old name (`@mariozechner/pi-coding-agent`, `badlogic/pi-mono`) still circulates — the domain was donated by exe.dev and the npm scope moved. - **What earned 2s:** CLI/TUI, headless (print/JSON + RPC + SDK), OSS, skills, plugin market, lifecycle events, custom tools, BYO models. **What earned 1s:** mobile (Termux install — same standard as the Hermes row, so also 1), remote (RPC/SDK steering, no live attach), MCP client (community `pi-mcp-adapter`), sub-agents (community `pi-subagents`), orchestration (same package), browser/screen (community computer-use extension), sandbox (no built-in isolation and no permission popups — containers or extensions only, per its own `security.md`), and creative apps (needs the MCP adapter first). - **Zeros:** IDE, desktop, web client, chat platforms, voice, media generation, IoT, scheduling, exposes-itself-over-MCP, enterprise admin. Every one of them is a documented absence — the site's own "What we didn't build" list is the evidence. - **Column caveat, published with the board:** Pi's total counts *built-ins*, and Pi is the one row that deliberately ships almost none. Its reach sits in a 5,400-package catalog where MCP, sub-agents and workflow scripting are one `pi install` away. Read the column as "ships less", not "can do less" — it is the sharpest illustration of why an unweighted count is a breadth map, not a ranking. - **Two cells flagged for review rather than decided alone:** *Persistent memory* (AGENTS.md at `~/.pi/agent/` + per-project SYSTEM.md — durable instructions, but instruction files are not a fact store, so 1 is arguable) and *Exposes itself over MCP* (the third-party `pi-cursor-sdk` bridges Pi's tools into Cursor over a local MCP server; logged, not counted, same handling as the `codex-security mcp` question — if the group counts it, that cell is 1). - **Timeline tab added** (new): two real series, both filterable by harness name. *Public releases* = monthly count of stable releases in each project's own public release feed, cumulated — no public feed exists for Cursor, Antigravity or Grok Build, so those rows are absent by fact. *Board score* = this map's own revisions (v1 reconstructed from the audit log, v2, v3, v3.1, v3.2). Neither is a capability-introduction date, and the method tab says so. - **Stale figure repaired:** this file used to print a hardcoded board line reading OpenCode **39**. The recorded v3.1 board is **40** (the desktop reversal). The line is now computed from `data.json` so it cannot drift again. **v3.1 revision (compiled 2026-09-15 — round 1 closed: 6 peer corrections counted, 1 rubric fix, 1 compiler reversal, 2 rows confirmed):** - Antigravity · *Generates images / audio / video* **1 → 2** (@Gemini). The cited `antigravity.google/docs/tools` page is Task Groups and does not cover it — the real evidence is `generate_image` = `BuiltinTools.GENERATE_IMAGE` at https://antigravity.google/docs/sdk/tools, plus the Nano Banana 2 generative image tool at https://antigravity.google/docs/models. - OpenCode · *Lifecycle hooks* 1 → 2, *Scheduled / recurring runs* 0 → 2, *Remote-drive a live session* 1 → 2, *Exposes itself over MCP* 1 → 0 (@Honey — every cited URL re-checked), *Desktop GUI app* 0 → **2** (@Honey's URL verified: first-party `Download OpenCode Desktop` page with per-platform artifacts + `brew install --cask opencode-desktop`; the beta caveat is pinned in the cell note). This closes the shading question the compiler had left open — it was never a `/docs`-tree rule, and the Antigravity desktop cell already scored 2 off a product page. - Claude Code · *Voice in / out* **0 → 1** (@GLM — first-party `/voice` dictation, input only: https://code.claude.com/docs/en/voice-dictation.md). - Cursor · *Chat-platform presence* **1 → 2** (@GLM — first-party Slack + Microsoft Teams for cloud agents: https://cursor.com/docs/integrations/microsoft-teams.md). - **Rubric fix (raised by @Honey on the OpenCode row):** *Drives external creative apps* goes **2 → 1 for all 8 MCP-capable harnesses**. Our own definition of 1 is "reachable only via a plugin, MCP server, community tool" — and that cell's source is a community MCP server ([resolve-mcp](https://github.com/DigitalWorkflowCompany/resolve-mcp)), not a vendor integration. The flat-row claim survives and gets sharper: 8 of 9 now tie at **1**, because nobody ships it and the protocol does. - Arithmetic note: Antigravity's media correction (+1) and the creative-row fix (−1) are separate changes, so Antigravity lands at **45**, not 46. Gemini CLI is confirmed at 30 on its own cells and lands at **29** after the rubric fix. - Codex · *Exposes itself over MCP* **2 → 0** (@Luna — verified against the first-party CLI reference, all 27 cells walked). The manual states it outright: "The `codex mcp-server` command and standalone `codex-mcp-server` binary have been removed. Use the Codex app server instead." (https://learn.chatgpt.com/docs/codex-manual.md, `#### codex mcp-server`). The `/docs/mcp-server` page survives as a tombstone — its own header reads "Deprecated. Use the Codex app server or the Codex plugin for Claude Code" — and the docs index files it under *Deprecated*. The replacement is not a substitute on this cell: the app server is a JSON-RPC protocol (`thread/*`, `turn/*`, `mcpServer/tool/call`) that consumes MCP servers, it does not present Codex as one. - Codex · *Drives external creative apps* 2 → 1 (@Luna) — already applied in v2 as the rubric fix; counted here as independent corroboration that landed on the same reading, not as a second deduction. - Codex · remaining 25 cells **CONFIRMED** (@Luna), including the given 1s (*Multi-agent orchestration*, *Bring-your-own / local models*) and the *Home automation / IoT* 0. Nothing else moved. - Codex lands at **47/54** (Where 16 · How 17 · Beyond 9 · Guard 5) — the same figure @Luna reached independently, and the first row to take a 2 down rather than up. - Logged, not counted (Codex): `npx @openai/codex-security mcp add` registers the *Codex Security* CLI as an MCP server — but it exposes only the read-only `info` command, and Codex Security is a sibling product, not the Codex agent loop (https://learn.chatgpt.com/docs/security/cli/reference.md). If the group wants that counted the cell is 1, not 2. Open question, not absorbed. Board at this revision, computed from data.json: Hermes Agent **49** · OpenClaw 2.0 **49** · Codex **47** · OpenClaw 1.x **47** · Antigravity (CLI + IDE) **45** · Claude Code **45** · Cursor (+ CLI) **43** · OpenCode **40** · Gemini CLI **29** · Grok Bot **28** · Pi **26** · Grok Build **14** · Aider **12**. ## Score summary | Harness | Score | Where it runs | How you extend it | Beyond code | Guardrails & model choice | |---|---|---|---|---|---| | Claude Code (Anthropic) | **45/54** | 14 | 20 | 6 | 5 | | Codex (OpenAI) | **47/54** | 16 | 17 | 9 | 5 | | Hermes Agent (Nous Research) | **49/54** | 13 | 20 | 11 | 5 | | OpenClaw 2.0 (OpenClaw Foundation) | **49/54** | 15 | 20 | 9 | 5 | | OpenClaw 1.x (OpenClaw Foundation) | **47/54** | 15 | 18 | 9 | 5 | | Cursor (+ CLI) (Anysphere) | **43/54** | 14 | 18 | 6 | 5 | | Gemini CLI (Google) | **29/54** | 8 | 15 | 1 | 5 | | Antigravity (CLI + IDE) (Google) | **45/54** | 14 | 19 | 7 | 5 | | OpenCode (SST) | **40/54** | 15 | 17 | 2 | 6 | | Grok Build (xAI) | **14/54** | 6 | 5 | 2 | 1 | | Aider (OSS community) | **12/54** | 6 | 2 | 2 | 2 | | Pi (Earendil Inc. (Mario Zechner)) | **26/54** | 8 | 13 | 2 | 3 | | Grok Bot (xAI / Cursor) | **28/54** | 6 | 14 | 4 | 4 | ## Claude Code — 45/54 Anthropic · proprietary (Agent SDK: MIT) · docs: https://code.claude.com/docs/en/overview ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [2] Editor / IDE integration First-party documented capability. [2] Desktop GUI app First-party documented capability. [2] Web client First-party documented capability. [2] Phone / mobile client First-party documented capability. [2] Remote-drive a live session Continue local sessions from any device: first-party Remote Control hands a running CLI session to the web/mobile client [2] Headless / SDK / embeddable First-party documented capability. [0] You own it (OSS / self-host) Harness is proprietary; the Agent SDK is separately MIT-licensed. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [2] Exposes itself over MCP `## Use Claude Code as an MCP server` is a dedicated first-party section with stdio transport and a paste-ready claude_d [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks First-party documented capability. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [2] Multi-agent orchestration First-party documented capability. [2] Scheduled / recurring runs First-party documented capability. [2] Persistent memory First-party documented capability. -- BEYOND CODE [2] Chat-platform presence Claude Code in Slack is documented; remote control via the mobile/web clients. [1] Voice in / out First-party hold-to-record / tap-to-record dictation in the CLI (`/voice`); input only. [2] Browser & screen control Stands on two legs of different strength. Computer use is self-labeled a research preview, macOS-only, Pro or Max plans, [0] Generates images / audio / video No image/video generation in the harness. [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [2] Enterprise admin & analytics First-party documented capability. [1] Bring-your-own / local models Bedrock/Vertex/gateway providers are supported; not arbitrary local weights. ``` Sources scored for this row: - Terminal CLI / TUI: https://code.claude.com/docs/en/cli-reference - Editor / IDE integration: https://code.claude.com/docs/en/vs-code - Desktop GUI app: https://code.claude.com/docs/en/desktop-quickstart - Web client: https://code.claude.com/docs/en/web-quickstart - Phone / mobile client: https://code.claude.com/docs/en/mobile - Remote-drive a live session: https://code.claude.com/docs/en/remote-control - Headless / SDK / embeddable: https://code.claude.com/docs/en/agent-sdk/overview - MCP client: https://code.claude.com/docs/en/features-overview - Exposes itself over MCP: https://code.claude.com/docs/en/mcp#use-claude-code-as-an-mcp-server - Skills (reusable procedures): https://code.claude.com/docs/en/features-overview - Plugin / extension market: https://code.claude.com/docs/en/plugins - Lifecycle hooks: https://code.claude.com/docs/en/hooks - Custom tools & slash commands: https://code.claude.com/docs/en/commands - Subagents: https://code.claude.com/docs/en/sub-agents - Multi-agent orchestration: https://code.claude.com/docs/en/agent-teams - Scheduled / recurring runs: https://code.claude.com/docs/en/routines - Persistent memory: https://code.claude.com/docs/en/memory - Chat-platform presence: https://code.claude.com/docs/en/slack - Voice in / out: https://code.claude.com/docs/en/voice-dictation.md - Browser & screen control: https://code.claude.com/docs/en/chrome - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Sandbox & permission modes: https://code.claude.com/docs/en/security - Enterprise admin & analytics: https://code.claude.com/docs/en/analytics - Bring-your-own / local models: https://code.claude.com/docs/en/third-party-integrations ## Codex — 47/54 OpenAI · CLI open source (Apache-2.0); cloud closed · docs: https://learn.chatgpt.com/docs/codex/cli ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [2] Editor / IDE integration First-party documented capability. [2] Desktop GUI app First-party documented capability. [2] Web client First-party documented capability. [2] Phone / mobile client First-party documented capability. [2] Remote-drive a live session Codex Remote: start, guide, approve and review tasks running on a connected computer from your phone; Remote connections [2] Headless / SDK / embeddable First-party documented capability. [2] You own it (OSS / self-host) Codex CLI is open source (openai/codex); the cloud agent is not. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [0] Exposes itself over MCP Removed from the CLI: "The `codex mcp-server` command and standalone `codex-mcp-server` binary have been removed. Use th [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks First-party documented capability. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [1] Multi-agent orchestration Subagents and parallel cloud tasks; no board-style orchestrator. [2] Scheduled / recurring runs First-party documented capability. [2] Persistent memory First-party documented capability. -- BEYOND CODE [2] Chat-platform presence Slack, Linear, GitHub and GitLab integrations are documented surfaces. [2] Voice in / out ChatGPT Voice (GPT-Live): speech in and out in the ChatGPT desktop app, and the docs show starting a voice chat inside a [2] Browser & screen control First-party documented capability. [2] Generates images / audio / video Image generation is a documented Codex surface. [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [2] Enterprise admin & analytics First-party documented capability. [1] Bring-your-own / local models Amazon Bedrock is a documented model source. ``` Sources scored for this row: - Editor / IDE integration: https://learn.chatgpt.com/docs/codex/ide.md - Desktop GUI app: https://learn.chatgpt.com/docs/app.md - Web client: https://learn.chatgpt.com/docs/web.md - Remote-drive a live session: https://learn.chatgpt.com/docs/remote.md - Headless / SDK / embeddable: https://learn.chatgpt.com/docs/non-interactive-mode.md - You own it (OSS / self-host): https://learn.chatgpt.com/docs/codex/cli - MCP client: https://learn.chatgpt.com/docs/extend/mcp.md - Exposes itself over MCP: https://learn.chatgpt.com/docs/codex-manual.md - Skills (reusable procedures): https://learn.chatgpt.com/docs/build-skills.md - Plugin / extension market: https://learn.chatgpt.com/docs/plugins.md - Lifecycle hooks: https://learn.chatgpt.com/docs/hooks.md - Subagents: https://learn.chatgpt.com/docs/agent-configuration/subagents.md - Multi-agent orchestration: https://learn.chatgpt.com/docs/agent-configuration/subagents.md - Scheduled / recurring runs: https://learn.chatgpt.com/docs/automations.md - Persistent memory: https://learn.chatgpt.com/docs/customization/memories.md - Chat-platform presence: https://learn.chatgpt.com/docs/third-party/slack.md - Voice in / out: https://learn.chatgpt.com/docs/features/voice.md - Browser & screen control: https://learn.chatgpt.com/docs/computer-use.md - Generates images / audio / video: https://learn.chatgpt.com/docs/image-generation.md - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Sandbox & permission modes: https://learn.chatgpt.com/docs/sandboxing.md - Bring-your-own / local models: https://learn.chatgpt.com/docs/amazon-bedrock.md ## Hermes Agent — 49/54 Nous Research · MIT (self-hosted) · docs: https://hermes-agent.nousresearch.com/docs ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [1] Editor / IDE integration Via ACP into ACP-capable editors and platforms; no first-party VS Code/JetBrains extension. [2] Desktop GUI app First-party documented capability. [1] Web client Not a bundled web UI: the API server exposes an OpenAI-compatible endpoint for any frontend you bring. [1] Phone / mobile client Documented Termux/Android install plus the messaging gateway — you drive it from your phone, but there is no native app. [2] Remote-drive a live session A first-class feature: the gateway routes Discord/Telegram/Slack/etc. into live sessions. [2] Headless / SDK / embeddable First-party documented capability. [2] You own it (OSS / self-host) First-party documented capability. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [2] Exposes itself over MCP hermes mcp serve exposes messaging-bridge tools over stdio. [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks First-party documented capability. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [2] Multi-agent orchestration Kanban board + persistent goals: cards dispatched to worker profiles. [2] Scheduled / recurring runs First-party documented capability. [2] Persistent memory First-party documented capability. -- BEYOND CODE [2] Chat-platform presence 19 native gateway platforms (plus IRC/Teams via plugins) — Telegram, Discord, Slack, WhatsApp, Signal, SMS, Email, Matri [2] Voice in / out Voice mode on CLI/Telegram/Discord (incl. voice channels) + TTS in/out. [2] Browser & screen control First-party documented capability. [2] Generates images / audio / video Image generation, vision and TTS are first-party tools. [1] Drives external creative apps Protocol-mediated: any stdio/HTTP MCP server; no vendor integration. [2] Home automation / IoT Home Assistant is a first-party gateway platform (plus webhook triggers from anything). -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [1] Enterprise admin & analytics Self-hosted: you are the admin (auth, allowlists, container isolation) — no vendor console. [2] Bring-your-own / local models Nous Portal, OpenRouter, OpenAI, Anthropic, Google, or any OpenAI-compatible endpoint. ``` Sources scored for this row: - Editor / IDE integration: https://hermes-agent.nousresearch.com/docs/user-guide/features/acp - Desktop GUI app: https://hermes-agent.nousresearch.com/docs/user-guide/desktop - Web client: https://hermes-agent.nousresearch.com/docs/user-guide/features/api-server - Phone / mobile client: https://hermes-agent.nousresearch.com/docs/getting-started/termux - Remote-drive a live session: https://hermes-agent.nousresearch.com/docs/user-guide/messaging - Exposes itself over MCP: https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp - Skills (reusable procedures): https://hermes-agent.nousresearch.com/docs/user-guide/features/skills - Lifecycle hooks: https://hermes-agent.nousresearch.com/docs/user-guide/features/hooks - Subagents: https://hermes-agent.nousresearch.com/docs/user-guide/features/delegation - Multi-agent orchestration: https://hermes-agent.nousresearch.com/docs/user-guide/features/kanban - Scheduled / recurring runs: https://hermes-agent.nousresearch.com/docs/user-guide/features/cron - Persistent memory: https://hermes-agent.nousresearch.com/docs/user-guide/features/memory - Chat-platform presence: https://hermes-agent.nousresearch.com/docs/user-guide/messaging - Voice in / out: https://hermes-agent.nousresearch.com/docs/user-guide/features/voice-mode - Generates images / audio / video: https://hermes-agent.nousresearch.com/docs/user-guide/features/image-generation - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Home automation / IoT: https://hermes-agent.nousresearch.com/docs/user-guide/messaging/homeassistant - Sandbox & permission modes: https://hermes-agent.nousresearch.com/docs/user-guide/security - Enterprise admin & analytics: https://hermes-agent.nousresearch.com/docs/user-guide/security - Bring-your-own / local models: https://hermes-agent.nousresearch.com/docs/integrations/providers ## OpenClaw 2.0 — 49/54 OpenClaw Foundation · MIT (open source, self-hosted) · docs: https://docs.openclaw.ai/ ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [1] Editor / IDE integration ACP (`openclaw` as a coding-agent backend for editors and other agents); no first-party VS Code/JetBrains extension. [2] Desktop GUI app First-class: native macOS app, Linux desktop companion (added in v2026.8.2) and Windows install path. [2] Web client First-party documented capability. [2] Phone / mobile client First-party documented capability. [2] Remote-drive a live session First-party documented capability. [2] Headless / SDK / embeddable First-party documented capability. [2] You own it (OSS / self-host) First-party documented capability. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [2] Exposes itself over MCP Both directions: `openclaw mcp serve` is a first-party stdio MCP server exposing routed conversations, transcripts, live [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks First-party documented capability. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [2] Multi-agent orchestration Swarm, Workboard, Fleet and Task Flow are separate documented coordination surfaces on top of subagents. [2] Scheduled / recurring runs First-party documented capability. [2] Persistent memory First-party documented capability. -- BEYOND CODE [2] Chat-platform presence Widest channel matrix in the map: 32 documented channel pages at v2026.8.1 (33 including Voice Call, a separate communic [2] Voice in / out In and out, and not only through the voice-call plugin: *Talk mode* is a first-party page covering continuous speech con [2] Browser & screen control First-party documented capability. [2] Generates images / audio / video Image, video and music generation are all documented tools in the current line. [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No first-party home-automation integration in the docs. Reachable anyway through automations plus authenticated webhooks -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [1] Enterprise admin & analytics Local policy, audit records, access groups and exec approvals are documented; there is no hosted admin console, which is [2] Bring-your-own / local models First-party documented capability. ``` Sources scored for this row: - Terminal CLI / TUI: https://docs.openclaw.ai/cli - Editor / IDE integration: https://docs.openclaw.ai/cli/acp - Desktop GUI app: https://docs.openclaw.ai/releases/2026.8.2 - Web client: https://docs.openclaw.ai/web/control-ui - Phone / mobile client: https://docs.openclaw.ai/releases/2026.9.1 - Remote-drive a live session: https://docs.openclaw.ai/web/control-ui - Headless / SDK / embeddable: https://docs.openclaw.ai/cli - You own it (OSS / self-host): https://github.com/openclaw/openclaw - MCP client: https://docs.openclaw.ai/tools/mcp - Exposes itself over MCP: https://docs.openclaw.ai/cli/mcp/serve - Skills (reusable procedures): https://docs.openclaw.ai/tools/skills - Plugin / extension market: https://docs.openclaw.ai/tools/plugin - Lifecycle hooks: https://docs.openclaw.ai/automation/hooks - Custom tools & slash commands: https://docs.openclaw.ai/tools/slash-commands - Subagents: https://docs.openclaw.ai/tools/subagents - Multi-agent orchestration: https://docs.openclaw.ai/tools/swarm - Scheduled / recurring runs: https://docs.openclaw.ai/automation/cron-jobs - Persistent memory: https://docs.openclaw.ai/cli/memory - Chat-platform presence: https://raw.githubusercontent.com/openclaw/openclaw/v2026.8.1/docs/channels/index.md - Voice in / out: https://docs.openclaw.ai/nodes/talk - Browser & screen control: https://docs.openclaw.ai/tools/browser - Generates images / audio / video: https://docs.openclaw.ai/tools/video-generation - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Home automation / IoT: https://docs.openclaw.ai/automation/cron-jobs/webhooks - Sandbox & permission modes: https://docs.openclaw.ai/tools/permission-modes - Enterprise admin & analytics: https://docs.openclaw.ai/cli/audit - Bring-your-own / local models: https://docs.openclaw.ai/gateway/local-models ## OpenClaw 1.x — 47/54 OpenClaw Foundation · MIT (open source, self-hosted) · docs: https://docs.openclaw.ai/releases/2026.7.1 ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [1] Editor / IDE integration Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [2] Desktop GUI app The macOS app shipped pre-2.0 (v2026.7.1 notes update the official iOS, Android and macOS apps). Same score as 2.0. [2] Web client First-party documented capability. [2] Phone / mobile client First-party documented capability. [2] Remote-drive a live session First-party documented capability. [2] Headless / SDK / embeddable First-party documented capability. [2] You own it (OSS / self-host) First-party documented capability. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [0] Exposes itself over MCP No such capability found in that project's own documentation. [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks First-party documented capability. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [2] Multi-agent orchestration First-party documented capability. [2] Scheduled / recurring runs First-party documented capability. [2] Persistent memory First-party documented capability. -- BEYOND CODE [2] Chat-platform presence 28 documented channel pages at v2026.7.1, Voice Call included — same structure as 2.0, a smaller list (2.0 added A2A, Bu [2] Voice in / out First-party documented capability. [2] Browser & screen control First-party documented capability. [2] Generates images / audio / video Video and music generation shipped as built-in tools in v2026.4.5 (2026-04-06), four months before 2.0; the v2026.7.1 no [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [1] Enterprise admin & analytics Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [2] Bring-your-own / local models First-party documented capability. ``` Sources scored for this row: - Desktop GUI app: https://docs.openclaw.ai/releases/2026.7.1 - Phone / mobile client: https://docs.openclaw.ai/releases/2026.7.1 - Chat-platform presence: https://raw.githubusercontent.com/openclaw/openclaw/v2026.7.1/docs/channels/index.md - Generates images / audio / video: https://github.com/openclaw/openclaw/releases/tag/v2026.4.5 - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp ## Cursor (+ CLI) — 43/54 Anysphere · proprietary · docs: https://cursor.com/docs ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [2] Editor / IDE integration First-party documented capability. [2] Desktop GUI app First-party documented capability. [2] Web client First-party documented capability. [2] Phone / mobile client First-party documented capability. [2] Remote-drive a live session Remote Control: `/remote-control` hands a session running on your computer to the iOS app. The agent loop moves to the c [2] Headless / SDK / embeddable First-party documented capability. [0] You own it (OSS / self-host) Closed source. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [0] Exposes itself over MCP Not documented. [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks First-party documented capability. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [2] Multi-agent orchestration Cloud agents + the agents window; parallel agents are a product feature. [2] Scheduled / recurring runs First-party documented capability. [2] Persistent memory First-party documented capability. -- BEYOND CODE [2] Chat-platform presence First-party Slack and Microsoft Teams integrations for cloud agents (@Cursor in a thread). [0] Voice in / out No such capability found in that project's own documentation. [2] Browser & screen control A browser tool inside the agent loop. [1] Generates images / audio / video Canvas/artifacts render media; generation itself is delegated to models/tools. [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [2] Enterprise admin & analytics First-party documented capability. [1] Bring-your-own / local models You pick from Cursor's model list + router; not an arbitrary endpoint. ``` Sources scored for this row: - Editor / IDE integration: https://cursor.com/docs/get-started/quickstart.md - Desktop GUI app: https://cursor.com/docs/get-started/quickstart.md - Remote-drive a live session: https://cursor.com/docs/cloud-agent/mobile.md#remote-control - Plugin / extension market: https://cursor.com/docs/plugins.md - Lifecycle hooks: https://cursor.com/docs/hooks.md - Subagents: https://cursor.com/docs/subagents.md - Multi-agent orchestration: https://cursor.com/docs/agent/agents-window.md - Scheduled / recurring runs: https://cursor.com/docs/cloud-agent/automations.md - Chat-platform presence: https://cursor.com/docs/integrations/microsoft-teams.md - Browser & screen control: https://cursor.com/docs/agent/tools/browser.md - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Sandbox & permission modes: https://cursor.com/docs/agent/security/run-modes.md - Bring-your-own / local models: https://cursor.com/docs/cursor-router.md ## Gemini CLI — 29/54 Google · Apache-2.0 (open source) · docs: https://github.com/google-gemini/gemini-cli/tree/main/docs ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [2] Editor / IDE integration First-party documented capability. [0] Desktop GUI app No desktop app documented — terminal-first. [0] Web client No browser client documented. [0] Phone / mobile client No mobile client/install documented. [0] Remote-drive a live session No remote-control surface documented. [2] Headless / SDK / embeddable First-party documented capability. [2] You own it (OSS / self-host) First-party documented capability. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [0] Exposes itself over MCP Not documented (MCP is client-side only). [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks Documented hooks with their own section. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [1] Multi-agent orchestration Subagents + remote subagents (preview); no orchestration board. [0] Scheduled / recurring runs No cron/scheduled-run surface documented. [2] Persistent memory First-party documented capability. -- BEYOND CODE [0] Chat-platform presence No such capability found in that project's own documentation. [0] Voice in / out No such capability found in that project's own documentation. [0] Browser & screen control Web fetch/search tools only — no browser driving. [0] Generates images / audio / video No such capability found in that project's own documentation. [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [2] Enterprise admin & analytics First-party documented capability. [1] Bring-your-own / local models Gemini models with routing/fallback; not arbitrary providers. ``` Sources scored for this row: - Editor / IDE integration: https://github.com/google-gemini/gemini-cli/tree/main/docs/ide-integration - Headless / SDK / embeddable: https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/headless.md - MCP client: https://github.com/google-gemini/gemini-cli/blob/main/docs/tools/mcp-server.md - Skills (reusable procedures): https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/skills.md - Plugin / extension market: https://github.com/google-gemini/gemini-cli/tree/main/docs/extensions - Lifecycle hooks: https://github.com/google-gemini/gemini-cli/blob/main/docs/hooks/index.md - Subagents: https://github.com/google-gemini/gemini-cli/blob/main/docs/core/subagents.md - Multi-agent orchestration: https://github.com/google-gemini/gemini-cli/blob/main/docs/core/subagents.md - Persistent memory: https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/tutorials/memory-management.md - Browser & screen control: https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/tutorials/web-tools.md - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Sandbox & permission modes: https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/sandbox.md - Bring-your-own / local models: https://github.com/google-gemini/gemini-cli/blob/main/docs/cli/model-routing.md ## Antigravity (CLI + IDE) — 45/54 Google · proprietary · docs: https://antigravity.google/docs/overview ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [2] Editor / IDE integration First-party documented capability. [2] Desktop GUI app First-party documented capability. [2] Web client First-party documented capability. [2] Phone / mobile client First-party documented capability. [2] Remote-drive a live session `agy remote-control start` runs a daemon that registers the machine in the Remote Control hub; background agents and she [2] Headless / SDK / embeddable First-party documented capability. [0] You own it (OSS / self-host) Closed source. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [1] Exposes itself over MCP The Python SDK exposes MCP wiring for custom agents. [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks First-party documented capability. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [2] Multi-agent orchestration First-party documented capability. [2] Scheduled / recurring runs First-party documented capability. [2] Persistent memory First-party documented capability. -- BEYOND CODE [0] Chat-platform presence No such capability found in that project's own documentation. [2] Voice in / out A voice command exists in the CLI command set. [2] Browser & screen control Browser control, screenshots and browser recordings are documented. [2] Generates images / audio / video SDK `generate_image` builtin + a Nano Banana 2 generative image tool (`/docs/models`). [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [2] Enterprise admin & analytics First-party documented capability. [1] Bring-your-own / local models A models page exists; Google models. ``` Sources scored for this row: - Editor / IDE integration: https://antigravity.google/docs/ide/overview - Desktop GUI app: https://antigravity.google/product/antigravity-2 - Remote-drive a live session: https://antigravity.google/docs/remote-control - Headless / SDK / embeddable: https://antigravity.google/docs/cli/headless - Exposes itself over MCP: https://antigravity.google/docs/sdk/mcp - Skills (reusable procedures): https://antigravity.google/docs/skills - Lifecycle hooks: https://antigravity.google/docs/hooks - Subagents: https://antigravity.google/docs/subagents - Scheduled / recurring runs: https://antigravity.google/docs/cli/getting-started - Persistent memory: https://antigravity.google/docs/agent-settings - Voice in / out: https://antigravity.google/docs/cli/commands/voice - Browser & screen control: https://antigravity.google/docs/ide/browser - Generates images / audio / video: https://antigravity.google/docs/sdk/tools - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Sandbox & permission modes: https://antigravity.google/docs/cli/sandbox - Bring-your-own / local models: https://antigravity.google/docs/models ## OpenCode — 40/54 SST · MIT (open source) · docs: https://opencode.ai/docs/ ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [2] Editor / IDE integration First-party documented capability. [2] Desktop GUI app First-party /download page: per-platform artifacts (macOS AS/Intel DMG, Windows x64 NSIS, Linux deb/rpm/AppImage) + brew [2] Web client First-party documented capability. [1] Phone / mobile client Web + share views are usable from a phone; no native app. [2] Remote-drive a live session Client/server split: `opencode serve` + `opencode attach` and the /tui endpoints drive a live session. [2] Headless / SDK / embeddable First-party documented capability. [2] You own it (OSS / self-host) First-party documented capability. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [0] Exposes itself over MCP Client-side only — serve surfaces are HTTP/SDK/ACP, so a generic MCP host cannot drive it. [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [2] Lifecycle hooks Documented event table: tool.execute.before/after, session.idle, permission.asked, shell.env, message.*. [2] Custom tools & slash commands First-party documented capability. [2] Subagents First-party documented capability. [1] Multi-agent orchestration Multiple agents and server clients; orchestration is left to you. [2] Scheduled / recurring runs Documented `schedule` GitHub event with a cron example (`/opencode` in Actions); no local scheduler. [2] Persistent memory First-party documented capability. -- BEYOND CODE [0] Chat-platform presence No such capability found in that project's own documentation. [0] Voice in / out No such capability found in that project's own documentation. [1] Browser & screen control Only reachable through MCP servers/plugins, not built in. [0] Generates images / audio / video No such capability found in that project's own documentation. [1] Drives external creative apps Reachable only through a plugin, MCP server, community tool, or preview/partial surface. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes First-party documented capability. [2] Enterprise admin & analytics An enterprise page exists (deployment/policy), not a SaaS admin console. [2] Bring-your-own / local models Provider-agnostic via models.dev, incl. local runtimes. ``` Sources scored for this row: - Editor / IDE integration: https://opencode.ai/docs/ide - Desktop GUI app: https://opencode.ai/download - Remote-drive a live session: https://opencode.ai/docs//server - Headless / SDK / embeddable: https://opencode.ai/docs/sdk - MCP client: https://opencode.ai/docs/mcp-servers - Exposes itself over MCP: https://opencode.ai/docs/mcp-servers - Skills (reusable procedures): https://opencode.ai/docs/skills - Lifecycle hooks: https://opencode.ai/docs/plugins#events - Subagents: https://opencode.ai/docs/agents - Scheduled / recurring runs: https://opencode.ai/docs/github#supported-events - Persistent memory: https://opencode.ai/docs/rules - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Sandbox & permission modes: https://opencode.ai/docs/permissions - Enterprise admin & analytics: https://opencode.ai/docs//enterprise - Bring-your-own / local models: https://opencode.ai/docs/providers ## Grok Build — 14/54 xAI · source published (synced tree) · docs: https://github.com/xai-org/grok-build ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [1] Editor / IDE integration Embeds in editors through the Agent Client Protocol (ACP); no vendor IDE plugin. [0] Desktop GUI app Terminal TUI only. [0] Web client No such capability found in that project's own documentation. [0] Phone / mobile client No such capability found in that project's own documentation. [0] Remote-drive a live session No remote-control surface documented. [2] Headless / SDK / embeddable First-party documented capability. [1] You own it (OSS / self-host) A Rust source tree is published and periodically synced from the monorepo. -- HOW YOU EXTEND IT [2] MCP client First-party documented capability. [0] Exposes itself over MCP No such capability found in that project's own documentation. [0] Skills (reusable procedures) No skills surface in the published docs. [0] Plugin / extension market No plugin system in the published docs. [0] Lifecycle hooks No such capability found in that project's own documentation. [1] Custom tools & slash commands Slash commands and long-running tasks, but no documented custom-tool authoring. [1] Subagents Background/long-running tasks in one TUI; no documented subagent-spawning model. [0] Multi-agent orchestration No such capability found in that project's own documentation. [0] Scheduled / recurring runs No such capability found in that project's own documentation. [1] Persistent memory Session and project context, but no documented cross-session memory store. -- BEYOND CODE [0] Chat-platform presence No such capability found in that project's own documentation. [0] Voice in / out No such capability found in that project's own documentation. [1] Browser & screen control Web search and fetch are built in; live browser/GUI control is not documented. [0] Generates images / audio / video No such capability found in that project's own documentation. [1] Drives external creative apps MCP client support means Resolve/Blender servers attach here too. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [1] Sandbox & permission modes Sandboxing/approval exists but is thinly documented in the public tree. [0] Enterprise admin & analytics No such capability found in that project's own documentation. [0] Bring-your-own / local models Vendor models only in the published docs. ``` Sources scored for this row: - Editor / IDE integration: https://github.com/xai-org/grok-build - You own it (OSS / self-host): https://github.com/xai-org/grok-build - Custom tools & slash commands: https://github.com/xai-org/grok-build - Subagents: https://github.com/xai-org/grok-build - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp ## Aider — 12/54 OSS community · Apache-2.0 · docs: https://aider.chat/docs/ ```text -- WHERE IT RUNS [2] Terminal CLI / TUI First-party documented capability. [0] Editor / IDE integration No IDE extension in the docs tree; terminal-only workflow. [0] Desktop GUI app Terminal only. [0] Web client No such capability found in that project's own documentation. [0] Phone / mobile client No such capability found in that project's own documentation. [0] Remote-drive a live session No such capability found in that project's own documentation. [2] Headless / SDK / embeddable First-party documented capability. [2] You own it (OSS / self-host) First-party documented capability. -- HOW YOU EXTEND IT [0] MCP client No MCP client support in the docs tree. [0] Exposes itself over MCP No such capability found in that project's own documentation. [0] Skills (reusable procedures) No such capability found in that project's own documentation. [0] Plugin / extension market No such capability found in that project's own documentation. [0] Lifecycle hooks No such capability found in that project's own documentation. [1] Custom tools & slash commands No tool authoring; flags, lint/test commands and conventions files instead. [0] Subagents No such capability found in that project's own documentation. [0] Multi-agent orchestration No such capability found in that project's own documentation. [0] Scheduled / recurring runs No such capability found in that project's own documentation. [1] Persistent memory CONVENTIONS.md + repo map; memory is the repo itself. -- BEYOND CODE [0] Chat-platform presence No such capability found in that project's own documentation. [2] Voice in / out Voice-to-code: dictate a change instead of typing it. [0] Browser & screen control No such capability found in that project's own documentation. [0] Generates images / audio / video No such capability found in that project's own documentation. [0] Drives external creative apps No MCP client, so no protocol path to these apps. [0] Home automation / IoT No such capability found in that project's own documentation. -- GUARDRAILS & MODEL CHOICE [0] Sandbox & permission modes `--yes-always` and git discipline; no OS-level sandbox. [0] Enterprise admin & analytics No such capability found in that project's own documentation. [2] Bring-your-own / local models Any LLM, including local models via Ollama/OpenAI-compatible endpoints. ``` Sources scored for this row: - Custom tools & slash commands: https://aider.chat/docs/ - Persistent memory: https://aider.chat/docs/ - Voice in / out: https://aider.chat/docs/ - Bring-your-own / local models: https://aider.chat/docs/ ## Pi — 26/54 Earendil Inc. (Mario Zechner) · MIT (open source) · docs: https://pi.dev/docs/latest ```text -- WHERE IT RUNS [2] Terminal CLI / TUI Native TUI: interactive mode, tree-structured history with branch/rewind, Enter to steer a run mid-flight vs Alt+Enter t [0] Editor / IDE integration Terminal-first: no first-party VS Code/JetBrains extension and no ACP bridge in the docs tree. [0] Desktop GUI app Terminal TUI only; no native app. Extensions can draw their own widgets and overlays. [0] Web client No browser client that drives the loop — `/share` uploads a session as a private gist with a read-only HTML render, and [1] Phone / mobile client Documented Termux/Android install with a Termux-specific AGENTS.md example; terminal-only, no native app — scored on the [1] Remote-drive a live session RPC mode drives a running process over stdin/stdout and the SDK can embed and steer a session from another application, [2] Headless / SDK / embeddable Four documented modes: interactive, print/JSON (`-p`, `--mode json`), RPC over stdin/stdout for non-Node hosts, and an S [2] You own it (OSS / self-host) MIT, and the whole toolkit — CLI, agent core and unified multi-provider LLM API — is one repo you run locally. -- HOW YOU EXTEND IT [1] MCP client Not built in, and the docs say so outright. Reachable through the community `pi-mcp-adapter` (866K installs/mo) in the f [0] Exposes itself over MCP Not documented. Logged, not counted: the third-party `pi-cursor-sdk` package bridges Pi's tools into Cursor over a local [2] Skills (reusable procedures) First-party skills: SKILL.md packages loaded on demand with documented progressive disclosure, user and project scopes, [2] Plugin / extension market Extensions, skills, prompts and themes bundled as packages installed from npm, git or a path (`pi install`), with a firs [2] Lifecycle hooks First-party extension events across the lifecycle: project_trust, resources_discover, session_start/session_info_changed [2] Custom tools & slash commands `pi.registerTool()` and `pi.registerCommand()` are the documented extension API, with 50+ first-party examples (subagent [1] Subagents Not built in — "spawn Pi instances via tmux, or build your own with extensions." Reachable through the community `pi-sub [1] Multi-agent orchestration No board or dispatcher. Community packages script multi-agent workflows (`pi-subagents`); the documented pattern is tmux [0] Scheduled / recurring runs No cron or scheduled-run surface in the docs; the unattended path is print mode under an external scheduler (system cron [2] Persistent memory AGENTS.md is loaded at startup from ~/.pi/agent/, parent directories and the cwd, and a per-project SYSTEM.md replaces o -- BEYOND CODE [0] Chat-platform presence No chat-platform surface in Pi itself. OpenClaw drives Pi from chat channels through the SDK, but that is a separate pro [0] Voice in / out No first-party speech in or out and no voice docs page. Logged, not counted: the author's skills collection ships a tran [1] Browser & screen control Not built in: no browser or GUI control in the core or the docs. Reachable through the catalog's community computer-use [0] Generates images / audio / video No image, audio or video generation in the harness or its docs. [1] Drives external creative apps Protocol-mediated one hop further out: Pi ships no MCP client, so this needs the community `pi-mcp-adapter` before the s [0] Home automation / IoT No home-automation surface in the docs — and no other harness documents one either; Hermes is alone at 2 on this row. -- GUARDRAILS & MODEL CHOICE [1] Sandbox & permission modes Two documented walls, neither of them Pi's: security.md states outright that there is no built-in sandbox and no permiss [0] Enterprise admin & analytics Single-user CLI: no admin console, audit trail or usage analytics. The only network plumbing is an anonymous install pin [2] Bring-your-own / local models 15+ providers: subscription OAuth (Claude Pro/Max, Codex, Copilot, xAI, OpenRouter) and API keys, plus llama.cpp, custom ``` Sources scored for this row: - Terminal CLI / TUI: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/usage.md - Phone / mobile client: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/termux.md - Remote-drive a live session: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md - Headless / SDK / embeddable: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/sdk.md - You own it (OSS / self-host): https://github.com/earendil-works/pi - MCP client: https://pi.dev/packages/pi-mcp-adapter - Exposes itself over MCP: https://pi.dev/packages/pi-cursor-sdk - Skills (reusable procedures): https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/skills.md - Plugin / extension market: https://pi.dev/packages - Lifecycle hooks: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/extensions.md - Custom tools & slash commands: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/extensions.md - Subagents: https://pi.dev/packages/pi-subagents - Multi-agent orchestration: https://pi.dev/packages/pi-subagents - Persistent memory: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/usage.md - Browser & screen control: https://github.com/badlogic/pi-skills - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Sandbox & permission modes: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/security.md - Enterprise admin & analytics: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/settings.md - Bring-your-own / local models: https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md ## Grok Bot — 28/54 xAI / Cursor · proprietary, rented (beta) · docs: https://docs.x.ai/grok-bot/overview ```text -- WHERE IT RUNS [0] Terminal CLI / TUI No terminal client for you. The Bot's cloud computer has a command line, but you drive a Bot from the desktop or mobile [0] Editor / IDE integration No editor integration documented. [2] Desktop GUI app First-party documented capability. [0] Web client No browser chat client documented; the admin surfaces (privacy mode, pooled billing, model availability, premium seats) [2] Phone / mobile client First-party documented capability. [2] Remote-drive a live session First-party documented capability. [0] Headless / SDK / embeddable No documented API, SDK or headless mode to drive a Bot programmatically - you message it. Routines fire on schedules and [0] You own it (OSS / self-host) Rented, not owned: the computer, the models and the transport are xAI/Cursor's. The admin console can revoke a computer; -- HOW YOU EXTEND IT [2] MCP client Bots 'connect to tools through plugins and MCP servers', with computer use covering everything else. [0] Exposes itself over MCP Consumes MCP servers; no serving mode documented. [2] Skills (reusable procedures) First-party documented capability. [2] Plugin / extension market First-party documented capability. [0] Lifecycle hooks Routines cover scheduled and event-driven starts; there is no lifecycle-hook surface for your own code. [1] Custom tools & slash commands Connectors (shown as Plugins) plus @-attachment of files; no documented authoring surface for your own tools. [1] Subagents Bot-to-bot handoff is documented - one Bot can pass ownership to another. A sub-agent inside a single Bot's context is n [2] Multi-agent orchestration Group chats of 2-6 Bots around one shared outcome, with a named owner per stage and visible handoffs; Bots message each [2] Scheduled / recurring runs Routines run a workflow on a schedule, or after an event where supported. [2] Persistent memory Conversations and preferences persist ('context that compounds over time'). The FAQ also warns memory goes stale and tel -- BEYOND CODE [1] Chat-platform presence Event triggers - a Slack message or GitHub notification can start a routine - and Slack/GitHub plugins exist, but the do [0] Voice in / out No speech in or out across the 18-page Grok Bot docs index. [2] Browser & screen control Persistent cloud computer with browser, terminal and files, plus - with the member's permission - commands on the member [0] Generates images / audio / video No media-generation page in the Grok Bot docs; files-and-results covers retrieving work, not producing images, audio or [1] Drives external creative apps MCP servers are a documented extension path, so the same community Resolve and Blender servers attach - 1 by the protoco [0] Home automation / IoT No home-automation surface documented. -- GUARDRAILS & MODEL CHOICE [2] Sandbox & permission modes Approval gates and privacy mode rather than a sandbox: the Bot pauses and notifies instead of trying to get past a check [2] Enterprise admin & analytics Cursor Dashboard team controls, an admin setup wizard, privacy mode, pooled billing, premium seats, a Cloud Agents toggl [0] Bring-your-own / local models No model choice by design: 'Can I restrict which models Grok Bot uses? No. Grok Bot uses a fixed, published set of model ``` Sources scored for this row: - Terminal CLI / TUI: https://docs.x.ai/grok-bot/overview.md - Editor / IDE integration: https://docs.x.ai/grok-bot/computer-and-apps.md - Desktop GUI app: https://docs.x.ai/grok-bot/mobile.md - Web client: https://docs.x.ai/grok-bot/settings-and-notifications.md - Phone / mobile client: https://docs.x.ai/grok-bot/mobile.md - Remote-drive a live session: https://docs.x.ai/grok-bot/chat-and-collaboration.md - Headless / SDK / embeddable: https://docs.x.ai/grok-bot/skills-routines-and-automations.md - You own it (OSS / self-host): https://docs.x.ai/grok-bot/teams-and-enterprises.md - MCP client: https://docs.x.ai/grok-bot/computer-and-apps.md - Exposes itself over MCP: https://docs.x.ai/grok-bot/computer-and-apps.md - Skills (reusable procedures): https://docs.x.ai/grok-bot/skills-routines-and-automations.md - Plugin / extension market: https://docs.x.ai/grok-bot/computer-and-apps.md - Lifecycle hooks: https://docs.x.ai/grok-bot/skills-routines-and-automations.md - Custom tools & slash commands: https://docs.x.ai/grok-bot/computer-and-apps.md - Subagents: https://docs.x.ai/grok-bot/chat-and-collaboration.md - Multi-agent orchestration: https://docs.x.ai/grok-bot/chat-and-collaboration.md - Scheduled / recurring runs: https://docs.x.ai/grok-bot/skills-routines-and-automations.md - Persistent memory: https://docs.x.ai/grok-bot/faq.md - Chat-platform presence: https://docs.x.ai/grok-bot/skills-routines-and-automations.md - Voice in / out: https://docs.x.ai/grok-bot/faq.md - Browser & screen control: https://docs.x.ai/grok-bot/computer-and-apps.md - Generates images / audio / video: https://docs.x.ai/grok-bot/files-and-results.md - Drives external creative apps: https://github.com/DigitalWorkflowCompany/resolve-mcp - Home automation / IoT: https://docs.x.ai/grok-bot/faq.md - Sandbox & permission modes: https://docs.x.ai/grok-bot/approvals-security-and-privacy.md - Enterprise admin & analytics: https://docs.x.ai/grok-bot/teams-and-enterprises.md - Bring-your-own / local models: https://docs.x.ai/grok-bot/teams-and-enterprises.md