MCP Setup
Use the endpoint and authentication your deployment supports. This guide describes API-key connections.
Access comes before configuration
You need an existing account, an accessible project, and an API key you are permitted to use. Confirm the server address with the installation administrator. A public documentation page does not establish that a server is reachable from your client.
OpenClaw uses the existing private integration. Other clients need a reachable deployment and compatible transport and authentication. Do not assume that a client can use a configuration just because it supports MCP.
Transport and endpoint
Streamable HTTP uses /api/mcp/stream. The existing SSE integration uses /api/mcp. Both authenticate an API key with Authorization: Bearer <api-key>. Use HTTPS on the deployment host; the browser login cookie is not an API key.
The connection values below are not a universal client configuration file. Put them in the fields required by your client and choose the matching transport. Keep the real key in its secret or credential settings, not in a URL.
Choose scope, then verify
In Settings → API Keys, use the workspace/project scope and read or read-write mode appropriate to the task. The key cannot grant more access than its owning account. Never broaden a key merely to make an example work.
After connecting, list the available tools, select your project and ask for existing context. Test reading before asking the agent to save a note. If access is refused, check the project, key scope and client transport with the administrator; do not keep retrying with other credentials.
{
"streamableHttp": "/api/mcp/stream",
"legacySse": "/api/mcp",
"headers": {
"Authorization": "Bearer YOUR_API_KEY"
}
}Make your agent read past decisions
A connection does not make an agent remember automatically. Add an instruction to its rules file: read project context before working, then save new decisions with their reasons. Review what the agent writes. The instruction contains no API key.
Instruction for your agent
The agent can look up the project with list-projects. Replace the project line if you know its ID.
The agent will find the project you name.
Add it to: CLAUDE.md
## Nekomind memory (MCP server "nekomind")
Find the projectId of this repository's project with `list-projects()`.
- At the start of every session, before you plan or edit anything, call `session-context(projectId, topic)` with the projectId (topic: a few words about the task). Read what it returns and follow the decisions in it; if the task would contradict one, say so and ask first.
- Before you change or remove a rule that exists there, read its full page with `get-document(documentId)` (every digest item has a documentId and a nodeId; pass the documentId here).
- After you make or learn a decision, record it with `add_memory(projectId, content)`: the first line is the title, then why (the reason and the alternatives you rejected). If it changes an existing memory, use `update_memory(projectId, nodeId, content)` with that item's nodeId instead of adding a duplicate.
- Write a graph, not a list: one thing (a person, a decision, an idea, a document) is one node, and every new note links to what it is about. Link in the text with a line such as `supports:: [[Retry policy]]` (any type, then two colons and the exact title), or write several nodes and their edges at once with `upsert-graph(projectId, nodes, edges)`. Reuse the edge types the project already has; never leave a new note unlinked.
- Never put secrets or personal data in a memory.