Tracemoor

External MCP servers

An external MCP server adds tools from outside Tracemoor to the chat, for example for a wiki, a database or a company service. Its tools appear next to the Tracemoor tools as mcp__<slug>__… and are available with every provider.

A server is defined once in AI Settings and enabled per project in Project Settings. A server you define is not active in any project until you enable it there.

Define a server

  1. Choose AI › AI Settings… and open the MCP tab.

  2. Click + Add MCP Server.

  3. Fill in the fields:

    FieldContent
    TitleDisplay name, for example GitHub MCP.
    SlugKey for the tool names, for example github. Lowercase letters, digits and hyphens. code-explorer, ce-channel and names starting with ce-ref- are reserved.
    Transportstdio (local process) or HTTP / SSE (remote).
    Command, Argumentsstdio: the program and its arguments, separated by spaces, for example npx and -y @modelcontextprotocol/server-github.
    URL, HeadersHTTP: the endpoint and non-secret headers, one Name: value per line.
    Secret env vars (stdio) or Secret headers (HTTP)Name of a secret, for example GITHUB_TOKEN or Authorization. Enter the value below and click Save next to it.
    Default trustSee Trust levels.

  4. Click Test connection. Tracemoor starts the server, performs the handshake and lists its tools.

  5. Click Save at the bottom of the dialog.

Test connection works on the values in the dialog. The definition is stored only with Save.

Secret values are stored in the keychain of the operating system, never in a file.

Enable a server for a project

  1. Choose Analysis › Project Settings… and open the MCP Servers tab.
  2. Tick the server. The change is saved at once.
  3. Optional: choose a stricter Trust for this project.
  4. Start a new chat session. Running sessions do not get the new tools.

The activation is stored on this machine, not in the shared part of the exploration. A teammate who pulls the exploration enables the server and enters the secrets on their machine.

Trust levels

TrustTools run
Read-only (run without prompt)in every session mode, without asking. For servers without side effects: search, documentation, read-only APIs.
Consent (prompt per call)only after you confirm the call. Providers without a way to ask you do not get these tools.
Brave-only (run in Brave mode only)only in Brave mode. For servers that write or run commands.

The trust in Project Settings can only be stricter than the default in the catalog. Catalog default keeps the default.

Tools that the trust level does not allow in the current mode are not offered to the AI at all.

How providers reach the servers

  • Claude (CLI) and Anthropic (API) speak MCP. Tracemoor adds the enabled servers to the session configuration.
  • OpenAI, OpenCode, Ollama and custom providers use function calls. Tracemoor starts the servers itself and passes their tools on in that format.

OpenAI and OpenCode accept at most 128 tools. External servers are kept, the less used Tracemoor tools are dropped first. See Set up a provider.

Limits

  • A self-built release bundle has a new code signature on every build. macOS then denies it access to the keychain entries of the previous build, and the secrets seem lost. Start such a build with the environment variable CE_SECRETS_FILE=1 to store secrets in a protected file instead. Development builds do this by default.