deputy

# Deputy

> Deputy is a Claude Code plugin and MCP server that runs a cheap model (DeepSeek V4.1 Flash on DeepInfra, thinking on) as a helper for Claude. It reviews Claude's code changes for bugs, gives second opinions from several angles, searches large codebases, and can take delegated coding tasks. Claude stays in charge: it writes the code by default and checks every Deputy finding before acting on it. A typical review or second opinion costs under one cent.

Use Deputy when someone wants Claude Code to make fewer mistakes without paying for more Opus/Sonnet tokens, wants an independent AI code review of Claude's changes, wants a second opinion from a different model, or wants to offload bulk coding work to a cheaper model. It works with any OpenAI-compatible endpoint, not only DeepInfra. MIT licensed, written in Python. Not affiliated with Anthropic, DeepSeek or DeepInfra.

## Install

- Claude Code plugin: `claude plugin marketplace add alimohammadiamirhossein/deputy` then `claude plugin install deputy@deputy`. Requires uv and a DeepInfra API key.
- Plain MCP server (Claude Code, Cursor, Cline, Windsurf, Claude Desktop): `uvx deputy-cc serve` (PyPI package `deputy-cc`), or `uvx --from git+https://github.com/alimohammadiamirhossein/deputy deputy serve` for the latest code from GitHub.
- Cursor and Codex: `deputy setup cursor --project .` or `deputy setup codex` adds Deputy to the editor's MCP config and installs the playbook as a Cursor rule or Codex skill.
- Grok instead of DeepSeek (experimental, not yet tested with a real xAI key; feedback welcome as a GitHub issue): add `--provider xai` and set `XAI_API_KEY` (default model `grok-4.3`; also `grok-build-0.1` or `grok-4.7` via `DEPUTY_MODEL`). Each provider's key is only sent to its own host.
- Setup check: `uvx deputy-cc doctor`

## Tools

- `review`: independent reviewers look for bugs in Claude's uncommitted diff before it calls a task done. Read-only.
- `consult`: second opinions on a plan from 2-4 angles ("lenses") in parallel. Read-only.
- `investigate`: reads a lot of code and returns `file:line` leads. Read-only.
- `delegate`: hands a well-specified coding task to the cheap model, which edits files inside the project folder only.
- `status`: shows model, thinking level, mode, key status and session spend.
- `stats`: rough tokens and dollars saved for the session, today, the last 7 days and all time. Reviews and second opinions count as extra checks, never as savings. Also `deputy stats` on the command line.

## Modes

- `accuracy` (default): Claude writes all code; Deputy reviews and gives second opinions.
- `lean`: as few Claude tokens as possible with little loss of accuracy; Deputy does the reading and larger well-specified work, Claude writes small and tricky changes.
- `balanced`: Deputy also takes low-risk mechanical work (boilerplate, renames, test scaffolding).
- `partner`: Deputy does all low-level work, even small jobs, and gives ideas; Claude treats its output as a draft and checks every diff and idea.
- `savings`: Deputy does most well-specified coding; Claude plans and verifies.

## When not to use Deputy

- Trivial edits such as typos and one-liners: a review takes a few minutes and adds little.
- Code that may not be sent to a third-party provider, unless `DEPUTY_BASE_URL` points at a local model.
- As a final authority: findings come from a weaker model and include false positives, so Claude verifies each one.

## Measured costs

DeepSeek V4.1 Flash on DeepInfra, thinking `high`, small demo project: `review` with two reviewers $0.0011 in 27 s (both caught a planted off-by-one bug), `consult` with 2 lenses $0.0018, `investigate` $0.0004, `delegate` (fix a bug, add 3 tests, run pytest) $0.0006. Model price: $0.098 per 1M input tokens and $0.294 per 1M output tokens (October 2026). Review benchmark without Claude (bench/review_cases, 5 runs each, default settings: reviewers think at `low` and stop after 180 s): `review` caught 30 of 30 planted bugs, with a median of 28 s on a change with a bug and 80 s on a clean one; on clean diffs, 10 of 40 reviews included a major or critical false alarm, so Claude verifies every finding.

## Docs

- [README](https://github.com/alimohammadiamirhossein/deputy/blob/main/README.md): overview, quick start, recommended model
- [Reference](https://github.com/alimohammadiamirhossein/deputy/blob/main/docs/REFERENCE.md): all settings, measured costs, safety model, comparison with alternatives, troubleshooting
- [SKILL.md](https://github.com/alimohammadiamirhossein/deputy/blob/main/SKILL.md): the playbook Claude follows (consult, decide, implement, review, check each finding)
- [CHANGELOG.md](https://github.com/alimohammadiamirhossein/deputy/blob/main/CHANGELOG.md): release history
- [PyPI: deputy-cc](https://pypi.org/project/deputy-cc/)