deputy¶
Your frontier model plans. A cheaper model does the long reading. Code checks every answer.
deputy is an MCP server. Your coding agent (Claude Code, or any MCP client) hands it a read-heavy task, such as "summarise this 900-line document" or "what changed in the last ten commits, and why". deputy runs it on a cheaper model (a local llama.cpp or Ollama model, or any OpenAI-compatible API), checks the answer with code, and returns a short result. Your agent reads the answer, not the files.
When it pays¶
| Delegate | Do it yourself |
|---|---|
| Long reads: summaries, inventories, cross-references, facts with their lines | Lookups in a short file or one line (flagged small_input) |
| Big logs and test output | Reading code you are about to edit |
| Questions about git history | Searches that only list files, and anything that needs a shell, the network or credentials |
On 21 everyday tasks on two real repositories, run twice, a reading task saved about 4,800 to 5,600 frontier tokens (estimated) and every answer was correct; a lookup saved about nothing: see evidence.
Get started¶
curl -LsSf https://raw.githubusercontent.com/tommasomarchionni/deputy/main/install.sh | sh
deputy init # point it at your model server
deputy doctor # check that it answers
deputy install claude-code --apply # once: register the MCP server
deputy instructions --for claude --apply # in each project: teach the agent when to delegate
Then read Install and connect an agent.
Status¶
Alpha: it works end to end and is tested; expect breaking changes. Verified with Claude Code; GitHub Copilot, Cursor and Codex are configured by the installer but not yet verified. What is done and what is next: roadmap.