MCP Configuration

In a nutshell: MCP (Model Context Protocol) lets Agents call external tools through a unified protocol; MCP Servers are managed directly by Claw and mounted on demand when a session starts.

When to use MCP

Connect an Agent to databases, internal APIs, filesystem indexes, or monitoring systems.
Reuse open-source MCPs (e.g., GitHub MCP, Slack MCP).
Run multiple MCP Servers on the same Claw and pick the right one per session.

Adding a new MCP Server on a Claw

Claw Management → current host → MCP → Add.
Fill in serverKey (must be unique), name, and defaultConfig (including transport / command / args / env, etc.).
Save — the system immediately validates the connection.
In a new session's "MCP picker", check the Servers you want to mount.

Transport and defaultConfig fields

Supported transports:

stdio: local process communicates via stdin/stdout (most common).
sse: Server-Sent Events.
http: standard HTTP.

JSON format. Example (stdio):

json
{
  "transport": "stdio",
  "command": "npx",
  "args": ["-y", "@modelcontextprotocol/server-postgres"],
  "env": {
    "PG_URL": "postgres://user:pwd@localhost/db"
  }
}

There is also manageDoc: a Markdown block of operational notes, visible only to the person who configured this MCP.

When creating a session, the input area shows an "MCP picker" where you check the Servers to mount. Selections persist with the session — the same Servers remain mounted when you resume a conversation.

Trigger sessions skip MCP by default

Sessions started by a Trigger do not mount MCP by default (to avoid credential ambiguity). Declare it explicitly in the Trigger configuration if needed.

Security and credential boundaries

Do not hard-code user-level secrets in defaultConfig

defaultConfig is a shared local configuration on the Claw host. User-scoped secrets should be injected via environment variables or dedicated Connection configurations.

The full MCP configuration is not exposed to the frontend

The Client UI shows only a summary, protecting credential fields from leaking.

Deleting an MCP does not automatically clean up dependent sessions

Before deleting, confirm that no running session has it mounted.

When to choose Skill versus MCP

Skill suits "reusable workflow scripts" the Agent loads as instructions on demand.
MCP suits "external system tools" — databases, internal services, third-party APIs exposed as tools.
Both can coexist: Skill provides the steps, MCP provides the tools.

Continue reading