Connect Your AI Agent
Register the EKOS MCP server with Claude Code or any MCP client.
EKOS serves its knowledge through ekos mcp serve. The default transport is stdio, which is what Claude Code uses.
Claude Code
claude mcp add ekos -- ekos --config /path/to/ekos.toml mcp serve --workspace /path/to/workspace
A path-free registration works too, using an environment variable:
claude mcp add ekos --env EKOS_WORKSPACE=/path/to/workspace -- ekos mcp serve
For a project in the EKOS home, name the project instead of a path:
claude mcp add ekos -- ekos --project my-project mcp serve # --tenant acme for another tenant
Check it with claude mcp get ekos (it should show Connected) or /mcp inside a session. Restart Claude Code after changing a registration.
The setup wizard's Agents step does this registration for Claude Code, Claude Desktop and VS Code, backing up each config file first.
Other clients
| Client type | Transport | Flag |
|---|---|---|
| Claude Code, most desktop clients | stdio | (default) |
| Several clients sharing one server | raw NDJSON over TCP | --tcp 127.0.0.1:7331 |
| URL-only clients (VS Code / Copilot, Visual Studio, ChatGPT) | Streamable HTTP at POST /mcp |
--http 127.0.0.1:8080 |
Bind TCP and HTTP to loopback or a trusted network only, and require a token when exposing them: MCP Configuration.
Through the web console
When a team runs the web console, an agent does not need its own server: the console is an MCP
endpoint per project, Streamable HTTP at /mcp/<tenant>/<project>, with an API token created in
Admin → API tokens:
{"url": "https://ekos.example.com/mcp/acme/shop",
"headers": {"Authorization": "Bearer ekos_…"}}
The agent sees only the tools its role allows, limited to what its rules let it see, and every call is in the tenant's audit log. A token can never review or administer. See Deployment.