Skip to content
Browse documentation
cli

Connect a coding agent

Add Smoothie to Codex or Claude Code and check that the connection works.

1. Add Smoothie to your agent

Add one of these entries to your agent's configuration. Replace the executable and project paths with absolute paths on your machine. This setup reads retained Framework context; it does not grant tool execution.

Codex launch entry
[mcp_servers.smoothie]
command = "/absolute/path/to/smoothie"
args = ["mcp", "serve", "--project", "/absolute/my-product", "--framework"]
Claude Code project entry
{
  "mcpServers": {
    "smoothie": {
      "command": "/absolute/path/to/smoothie",
      "args": ["mcp", "serve", "--project", "/absolute/my-product", "--framework"]
    }
  }
}
  • Codex: merge into config.toml. Project-local config requires a trusted project.
  • Claude Code: merge into .mcp.json. Preserve existing servers and trust prompts.
  • Use absolute paths; GUI agents may not inherit PATH or expand ~.

Review an existing smoothie entry before replacing it. To use your own Markdown instead, see Context and search.

Optional: connect approved tools

Already approved external tools? Generate a host entry for it:

bash
smoothie mcp host-config --host codex --project /absolute/my-product \
  --approved github --format json

Copy data.snippet, not the full report. Review data.requiredEnvironment and the approval details first. Use --host claude for JSON instead of TOML. The generator requires at least one --approved selection; it writes no host file, starts nothing, and grants no new access.

For semantic search, follow Context and tool discovery. Its startup flags are not options of this generator.

2. Check the connection

Reconnect your agent and check its server status and tools/list. Try searching the selected docs, then reading a result. Follow the returned input schemas. Saving a config entry alone does not confirm a working connection.

If it fails, check the executable path, project root, source access, and stderr. Pass required environment variables to the agent process, not a prompt or committed config. Keep shell banners off stdio.

These examples cover local Codex/Claude stdio setup, not HTTP/SSE or every client/platform. A shell-capable agent can also use CLI JSON.