OpenClaw Native MCP Setup
Status: documented setup, not yet manually validated.
This runbook tests OpenClaw through its native CLI and MCP registry. Do not use
ollama launch for this validation path.
Goal
Verify that a real OpenClaw agent can use ai-memory-hub as a Streamable HTTP MCP server:
- Register ai-memory-hub in OpenClaw's native MCP registry.
- Probe the MCP server from OpenClaw.
- Ask OpenClaw to save a short memory through ai-memory-hub tools.
- Verify the saved memory directly through ai-memory-hub.
- Record issues before marking OpenClaw supported.
Current Research Notes
- OpenClaw installs as a Node-based CLI. Its README recommends Node 24 or
Node 22.19+,
npm install -g openclaw@latest, andopenclaw onboard --install-daemon. - OpenClaw's native CLI has an
openclaw mcpcommand family.openclaw mcp servemakes OpenClaw act as an MCP server. The otheropenclaw mcpsubcommands manage OpenClaw-owned outbound MCP server definitions. - OpenClaw docs say saved MCP definitions live under
mcp.servers, supporttransport: "streamable-http"for HTTP MCP servers, and can be probed withopenclaw mcp probeoropenclaw mcp doctor --probe. - OpenClaw exposes configured MCP servers as plugin-owned tools through the
bundle-mcpplugin. Sandbox/tool policy can hide those tools unlessbundle-mcp,group:plugins, or specific server tool globs are allowed.
Prerequisites
- ai-memory-hub running with MCP enabled at
http://127.0.0.1:8000/mcp/. - OpenClaw installed and onboarded.
- A working OpenClaw model/provider configuration.
- For first validation, keep ai-memory-hub bound to localhost and use OAuth
resource-server mode.
auth: noneis reserved for maintainer-only smoke tests and CI fixtures.
Install OpenClaw natively:
npm install -g openclaw@latest
openclaw onboard --install-daemon
openclaw gateway status
Start ai-memory-hub from this repo in whichever local mode you are validating. For example, use the checked-in PGVector stack:
cd examples/local-stack
docker compose up --build
Confirm ai-memory-hub readiness:
curl -fsS http://127.0.0.1:8000/ready
Register ai-memory-hub In OpenClaw
Add ai-memory-hub as a Streamable HTTP MCP server with OAuth:
openclaw mcp add ai-memory-hub-local --url http://127.0.0.1:8000/mcp --transport streamable-http --auth oauth
openclaw mcp login ai-memory-hub-local
openclaw mcp login ai-memory-hub-local --code <code>
Run the final command after browser approval, replacing <code> with the
authorization code returned by the approval flow.
Inspect the saved definition:
openclaw mcp show ai-memory-hub-local --json
openclaw mcp status ai-memory-hub-local --verbose
Probe the live server:
openclaw mcp probe ai-memory-hub-local --json
openclaw mcp doctor ai-memory-hub-local --probe
Expected probe result:
- OpenClaw connects to
http://127.0.0.1:8000/mcp. - The tool list includes ai-memory-hub memory tools.
- No auth, transport, or tool-filter diagnostics block usage.
Tool Policy Check
If OpenClaw can probe the server but the agent cannot see or call the tools, check tool policy and sandbox gates.
For a local first pass, use a tool profile that can expose plugin tools:
openclaw config get tools.profile
If sandboxing is enabled for the session, ensure MCP/plugin tools are allowed in
the sandbox tool policy. OpenClaw docs describe bundle-mcp, group:plugins,
and server-specific globs as the relevant allowlist entries for configured MCP
servers.
After changing OpenClaw config, run:
openclaw doctor
openclaw mcp doctor ai-memory-hub-local --probe
Manual Validation Prompt
Run OpenClaw natively:
openclaw agent --message "Use the configured ai-memory-hub MCP server. First call memory_validate, then memory_insert, then memory_retrieve. Save a short conversation with source openclaw, title OpenClaw Native MCP Test, and one user message: The OpenClaw native MCP validation phrase is coral-index. After saving, report the memory ID."
Expected behavior:
- OpenClaw discovers the configured ai-memory-hub MCP tools.
- OpenClaw validates the conversation before insert.
- OpenClaw inserts one memory with
source: "openclaw". - OpenClaw retrieves the returned memory ID.
- OpenClaw reports the memory ID in the final answer.
If OpenClaw uses server-prefixed tool names, the tools may appear with an
ai-memory-hub__ prefix. That is acceptable as long as the underlying
ai-memory-hub MCP tools are called.
Verify Outside OpenClaw
Use ai-memory-hub directly to confirm persistence:
curl -fsS http://127.0.0.1:8000/memory/search \
-H 'Content-Type: application/json' \
-d '{"query":"coral-index","source":"openclaw","top_k":5}'
Ask over the saved memory:
curl -fsS http://127.0.0.1:8000/memory/ask \
-H 'Content-Type: application/json' \
-d '{"question":"What is the OpenClaw native MCP validation phrase?","source":"openclaw","top_k":5}'
Expected direct verification:
- Search returns at least one result with
conversation.sourceset toopenclaw. - Ask returns an answer containing
coral-index. - Citations or provenance point back to the OpenClaw-inserted memory.
Bearer Auth Follow-Up
After localhost validation works, repeat with ai-memory-hub OAuth resource-server auth enabled.
Open questions to verify before documenting exact auth syntax:
- Whether OpenClaw stores static HTTP headers literally in
mcp.servers. - Whether OpenClaw supports secret references for MCP HTTP headers.
- Whether
openclaw mcp loginis useful only for OAuth servers, or can help with ai-memory-hub's bearer/API-key mode.
Until this is verified, do not commit bearer tokens into OpenClaw config. Prefer
OAuth resource-server validation bound to 127.0.0.1; reserve unauthenticated
mode for maintainer-only smoke tests.
Supported Status Checklist
Do not mark OpenClaw supported until these are complete:
- [x] Native setup documented.
- [ ]
openclaw mcp probe ai-memory-hub-local --jsonsucceeds. - [ ] OpenClaw agent can call
memory_validate. - [ ] OpenClaw agent can call
memory_insert. - [ ] OpenClaw agent can call
memory_retrieve. - [ ] Direct ai-memory-hub search finds the OpenClaw-inserted memory.
- [ ] Direct ai-memory-hub ask answers from the OpenClaw-inserted memory.
- [ ] Any ai-memory-hub issues discovered during manual validation are fixed.
- [ ] Bearer-token behavior is tested or explicitly documented as unsupported.
- [ ] A stable headless/scriptable OpenClaw command is confirmed before adding automated real-client coverage.
Known Risks
- OpenClaw has broad local tool access by design. Keep this validation on a local test machine and avoid exposing ai-memory-hub beyond localhost until auth is verified.
- OpenClaw tool policy can successfully register an MCP server while still
hiding its tools from an agent turn. Check
bundle-mcpand sandbox allowlists when tools do not appear. - The first supported claim should come after real manual validation, not after a successful config/probe command alone.
References
- OpenClaw README: https://github.com/openclaw/openclaw
- OpenClaw MCP CLI docs: https://docs.openclaw.ai/cli/mcp
- OpenClaw tool policy docs: https://docs.openclaw.ai/gateway/config-tools
- ai-memory-hub real-client plan:
real_client_mcp_smoke_plan.md