Agents and the MCP server
How to connect an agent to Sparrow through its MCP server, and set up the effort-to-model mapping for your harness.
The Sparrow MCP server runs on your machine and calls Sparrow's API as you, so an agent sees
and changes only your projects. From a checkout of the repository, in service-sparrow/:
export SPARROW_URL=https://<the dashboard's address>
uv run sparrow login
claude mcp add sparrow -e SPARROW_URL="$SPARROW_URL" -- uv --directory "$PWD" run sparrow mcp
Any MCP client can run uv run sparrow mcp the same way.
You also need to set up the effort-to-model mapping for your harness. The repository's root
AGENTS.md, in "Delegation by effort", names the models: low effort goes to Haiku 5.5, medium
to Sonnet 5.5, and high to Opus 5.5, which the effort-high subagent (the planner) runs. The
session itself can run a cheaper model, because it only orchestrates. For omp, the project
agents .omp/agents/effort-low.md, .omp/agents/effort-medium.md and
.omp/agents/effort-high.md each pin a model and point at that section, and omp's task tool
runs a piece with agent: effort-low, agent: effort-medium or agent: effort-high. To
change a model, edit the table in AGENTS.md and the harness files it lists, in one commit.
For Claude Code, the project subagents .claude/agents/effort-low.md,
.claude/agents/effort-medium.md and .claude/agents/effort-high.md each pin a model ID
(model: claude-haiku-5-5, model: claude-sonnet-5-5 and model: claude-opus-5-5) and
point at that section, and Claude Code delegates to them by name. Cursor reads the same
.claude/agents/ files, so it needs no files of its own; on Cursor, Claude models are
third-party models billed per token. Codex isn't set up yet, because its custom agents can't
choose the model provider. .github/scripts/test_agent_models.py fails when a harness file's
model differs from the table.