Let Claude Drive Grackle
Point an external agent — Claude Code, Copilot CLI, Codex CLI, any MCP client — at Grackle's MCP server, and it can drive Grackle itself: create tasks, spawn agents, search the knowledge graph, read logs. The orchestrator, one level up.
The endpoint
grackle serve exposes an MCP server at http://127.0.0.1:7435/mcp by default. It speaks streamable HTTP.
grackle serve
Loopback by default — the MCP server binds wherever the Grackle server binds (GRACKLE_HOST, 127.0.0.1 out of the box). grackle serve --allow-network binds 0.0.0.0 and opens it to the LAN; to reach it from another machine, do that or front it with a reverse proxy you control. (The standalone @grackle-ai/mcp binary is stricter — it refuses any non-loopback bind.)
Connect a client
An MCP client connects over HTTP. Most clients take a server entry like this:
{
"mcpServers": {
"grackle": {
"type": "http",
"url": "http://127.0.0.1:7435/mcp",
"headers": { "Authorization": "Bearer <your-api-key>" }
}
}
}
The API key is the one Grackle generated on first run (~/.grackle/api-key). Clients that speak OAuth can skip the static token: Grackle advertises its authorization server at /.well-known/oauth-protected-resource/mcp and the client negotiates from there.
What the agent can do
Once connected, Grackle's tools show up in the client namespaced mcp__grackle__<tool>. A few it will reach for:
| Tool | Does |
|---|---|
mcp__grackle__task_create | Create a task in a workspace |
mcp__grackle__session_spawn | Spawn an agent against an environment |
mcp__grackle__knowledge_search | Search the knowledge graph by concept |
So you can sit in Claude Code and say:
Create a task to add rate limiting to the auth service, then spawn an agent on it.
and it does — through Grackle, on a real environment, with its own credentials and its own entry in the log.
Next
- MCP Server — the full tool catalog and connection detail.
- Knowledge Graph — what the agents leave behind for the next one.
- Web UI — watch the work the driven agent kicks off.