mirror of
https://github.com/openclaw/openclaw.git
synced 2026-09-29 17:19:00 +08:00
Vincent KocandVincent Koc
2393e17f27
* docs: fix one-way and absolute links across cli, tools, gateway, and channels Closes the open `link`-kind audit findings filed against docs/cli/, docs/tools/, docs/gateway/ and docs/channels/. - Convert 21 absolute `https://docs.openclaw.ai/...` links to root-relative paths on protocol, clients, external-apps, embedding, protocol/transport, and zaloclawbot so local previews and versioned builds resolve them. - Add the missing reverse link for one-way Related entries, using the house bullet or `<Card>` shape each page already uses. - Link terms that were named but not linked: SecretRef and `gateway.trustedProxies` on sms, the meeting providers on transcripts, `openclaw models`/`openclaw agent` on infer, Talk on tts, `/tools/invoke` on the gateway index. - Give distinct link text to the two "macOS platform notes" links on gateway/troubleshooting. - Point the IRC workspace `.env` link at the section anchor rather than the Security index. - Add Related sections to pages that had none (sms, transcripts, promos, progress-card, onboard, access-groups, concepts/memory, operator-scopes). - Add the 28 zh-CN glossary sources the new list-item link labels require, each inserted beside a related existing term. No anchor targets change. `pnpm check:docs` is green. * docs: link devices and doctor inline on pairing, and trusted proxy auth from the security index Completes two findings that were only half-applied: r3-1324 asked for the sibling commands to be linked inline on /cli/pairing, and r3-1579 lists the security index among the four pages that should link back to /gateway/trusted-proxy-auth. * docs: match each Related list's existing separator style The added bullets used an em dash in lists whose existing entries separate the label and gloss with a hyphen. Normalized 11 files so each list stays internally consistent. * docs: correct four inaccurate Related-link descriptions from review ClawSweeper found that several glosses I wrote misdescribed their targets. Verified each against the destination page and corrected: - ask_user is for non-secret questions; its own reference says never to answer it with a credential, so the secrets-page gloss now points credential entry at the masked secrets tool (P1). - OpenClaw sandbox policy does not wrap ACP harness execution (tools/acp-agents/delivery.md), so the ACP backlink now names the separate boundary instead of implying coverage (P2). - Code Mode is a QuickJS-WASI JavaScript/TypeScript orchestration surface, not something built on the xAI code_execution tool documented on that page (P2). - openclaw qr emits a mobile-node bootstrap payload, not a channel DM pairing code, on both the pairing and devices pages (P2). Audited the other 100 descriptions for the same class of error and fixed six more: OpenShell is a managed sandbox backend rather than a surface the sandbox bounds (3 pages); a Cloud Workers gloss claimed sandbox settings apply when its managed workspace is explicitly not an OS sandbox; openclaw promos lists and claims promotional offers; the 1Password gloss overstated the plugin relation; and the progress-card bullets no longer assert undocumented dashboard rendering. * docs: correct three more link descriptions and audit every remaining gloss Addresses the second review round, and then re-checks the whole set rather than only the three reported. Reported: - openclaw attach launches Claude Code with a scoped Gateway MCP grant (cli/attach.md:2); it does not attach a terminal to a running session. Fixed on cli/mcp.md and cli/tui.md. - Bare openclaw setup opens an interactive system-agent chat and falls through to guided onboarding on a fresh system (cli/setup.md:12), so calling it 'the non-interactive setup command' was wrong. - Access groups have no pairing prerequisite: a group 'grants nothing by itself' and only matters where an allowlist references it. Dropped the invented sequencing claim. Found by auditing the remaining descriptions against each target's own summary: - openclaw worker is explicitly 'not a general-purpose command for manual worker registration' (cli/worker.md:13); reworded on cloud-workers and protocol/handshake. - /reference/rpc documents RPC adapters for external CLIs, not worker RPC method shapes; repointed to /gateway/protocol/rpc-methods. - stable-https-url gives a tailnet-only URL via Tailscale Serve, which the iOS and Android glosses omitted. - Tightened health, logs, sandbox, resume, infer, dashboard, diffs, prometheus, music-generation, subagents, pairing, audit, tools-invoke, configuration-reference, and the three goal glosses to match their pages instead of asserting relationships no page documents. One glossary source added for the new label. * docs: attribute goal reads and updates to the dedicated goal tools The session-tool page documents session discovery, messaging, lifecycle, and orchestration tools; goal operations use get_goal, create_goal, and update_goal instead (docs/tools/goal.md:132-149). I corrected the other three goal glosses last round and missed this one. --------- Co-authored-by: Vincent Koc <vincent@openclaw.org>
11 KiB
11 KiB
summary, read_when, title, sidebarTitle
| summary | read_when | title | sidebarTitle | |||
|---|---|---|---|---|---|---|
| Expose OpenClaw channel conversations over MCP and manage saved MCP server definitions |
|
MCP | MCP |
openclaw mcp has two jobs:
- run OpenClaw as an MCP server with
openclaw mcp serve - manage OpenClaw-managed outbound MCP server definitions with
list,show,status,doctor,probe,add,set,configure,tools,login,logout,reload, andunset
serve is OpenClaw acting as an MCP server. The other subcommands are OpenClaw acting as an MCP client-side registry for servers its own runtimes may consume later.
Use openclaw acp when OpenClaw should host a coding harness session itself and route that runtime through ACP.
Choose the right MCP path
| Goal | Use | Why |
|---|---|---|
| Let an external MCP client read/send OpenClaw channel conversations | openclaw mcp serve |
OpenClaw is the MCP server and exposes Gateway-backed conversations over stdio. |
| Save third-party MCP servers for OpenClaw-managed agent runs | openclaw mcp add, set, configure, tools, login |
OpenClaw is the MCP client-side registry and later projects those servers into eligible runtimes. |
| Check a saved server without running an agent turn | openclaw mcp status, doctor, probe |
status and doctor inspect config; probe opens a live MCP connection and lists capabilities. |
| Edit MCP config from a browser | Control UI /settings/mcp (/mcp alias) |
The page shows inventory, enablement, OAuth/filter summaries, command hints, and a scoped mcp editor. |
| Give Codex app-server a scoped native MCP server | mcp.servers.<name>.codex |
The codex block only affects Codex app-server thread projection and is stripped before native config handoff. |
| Run ACP-hosted harness sessions | openclaw acp and ACP Agents |
ACP bridge mode does not accept per-session MCP server injection; configure gateway/plugin bridges instead. |
MCP pages
This page is an index. openclaw mcp has six pages, one per reader job. Open
the page that matches your task.
| Page | Read it when |
|---|---|
| Run OpenClaw as an MCP server | An MCP client should read or send OpenClaw channel conversations through openclaw mcp serve. |
| Manage saved MCP servers | You are saving, inspecting, or approving third-party MCP servers for OpenClaw-managed runs. |
| JSON output shapes | You are scripting against status --json, doctor --json, or probe --json. |
| Transports and OAuth | You need a transport config field, or you are running the MCP OAuth login flow. |
| MCP in the Control UI | You want to edit or inspect MCP config from a browser. |
| MCP Apps | You are enabling or securing the MCP Apps host bridge. |
Where each section moved
Every anchor from the previous single-page version still resolves here, so an
existing link such as /cli/mcp#bridge-tools keeps working. Each entry points at
the page that now holds the content.
- OpenClaw as an MCP server
- When to use serve
- How it works
- Client spawns the bridge
- Bridge connects to Gateway
- Sessions become MCP conversations
- Live events queue
- Optional Claude push
- Important behavior
- Choose a client mode
- Generic MCP clients
- Claude Code
- What serve exposes
- Usage
- Local Gateway
- Remote Gateway (token)
- Remote Gateway (password)
- Verbose / Claude off
- Bridge tools
- conversations_list
- conversation_get
- messages_read
- attachments_fetch
- events_poll
- events_wait
- messages_send
- permissions_list_open
- permissions_respond
- Event model
- Claude channel notifications
- off
- on
- auto (default)
- MCP client config
- Options
- --url
- --token
- --token-file
- --password
- --password-file
- --claude-channel-mode
- -v, --verbose
- Security and trust boundary
- Testing
- Troubleshooting
- No conversations returned
- events_poll or events_wait misses older messages
- Claude notifications do not show up
- Approvals are missing
- OpenClaw as an MCP client registry
- Important behavior
- Codex tool approvals
- Saved MCP server definitions
- Common server recipes
- Filesystem
- Memory
- Local script
- Remote HTTP
- Desktop/CUA
- JSON output shapes
- status --json
- doctor --json
- probe --json
- Stdio transport
- SSE / HTTP transport
- SSE / HTTP transport
- OAuth workflow
- Save the server
- Start login
- Use the manual fallback when needed
- Check authorization
- Clear credentials
- Streamable HTTP transport
- Control UI
- MCP Apps
- Current limits
Related
- Connect MCP servers
- CLI reference
- Plugins
openclaw attach— launch Claude Code with a temporary session-scoped Gateway MCP grant