deputy

Deputy

Deputize cheap models. Claude stays the sheriff.

Deputy is a Claude Code plugin and MCP server that has a cheap model, DeepSeek V4.1 Flash, review Claude's code for bugs and give second opinions, for under a cent each. Claude keeps writing the code.

A replay of one benchmark case. Deputy's findings, time and cost, the repro and the test run are real.

claude plugin marketplace add alimohammadiamirhossein/deputy
claude plugin install deputy@deputy

Inside Claude Code the same thing is /plugin marketplace add alimohammadiamirhossein/deputy. You need uv and a DeepInfra API key.

What it costs (measured)

Run	Time	Cost
review: two reviewers check a change with a planted off-by-one bug (both caught it)	27 s	$0.0011
consult: second opinions from 2 angles	68 s	$0.0018
investigate: list a project's functions and their tests	18 s	$0.0004
delegate: fix a bug, add 3 tests, run pytest until green	46 s	$0.0006

Measured on a small demo project with thinking high. Reviews and second opinions now think at low by default, which makes them faster and cheaper (see the review benchmark below). On real projects, expect consult and review to stay under $0.01. Every result shows its real tokens and cost.

Benchmark: Opus with and without Deputy

Four small Python tasks with hidden tests Claude never sees (bench/), one run each on Claude Opus 5.5, October 2026. These runs predate the faster reviews: reviewers then thought at high with no time limit, and the prompt asked Claude to review every task.

Config	Hidden tests passed	Mean cost (Claude + Deputy)	Median time
Opus alone	4/4	$0.18	30 s
Opus + /deputy:review	4/4	$0.31 ($0.30 + $0.013)	3.7 min
Opus + /deputy:delegate	4/4	$0.19 ($0.19 + $0.0015)	76 s

What this shows, and what it doesn't:

• No accuracy gain to measure here. The tasks are too easy: Opus passed every hidden test on its own.
• Reviews cost Claude more than they cost Deputy. Deputy's share is about a cent, but Claude spends more checking the findings, so a reviewed task cost 68% more and took longer. One review took 24 minutes because a reviewer thought for 73k tokens. Reviewers now think at low and stop after 180 s, and Claude reviews only risky or large changes.
• The findings still changed the code. In 3 of 4 tasks Claude hardened the code beyond the spec: it fixed a rate limiter that a NaN timestamp would block forever, and added missing tests. It rejected one finding that contradicted the spec.
• Delegating saved nothing on tasks this small. Claude's own reading and checking cost about as much as writing the code itself.
• Small sample. One run per config. Full table: bench/results/2026-10-24-opus.

Does review catch bugs, and how often does it cry wolf?

Deputy's review on its own, without Claude, on 6 changes with a planted bug and 8 clean changes. Each planted bug has a clean twin, the same change done right. Five runs each with the default settings (bench/review_cases):

Diffs	Reviews	Flagged	Flagged major or critical	Mean cost	Median time	Slowest
Planted bug	30	30/30 caught	30/30	$0.0016	28 s	59 s
Clean	40	22/40	10/40	$0.0039	80 s	230 s

• Every planted bug was caught, at major or critical severity, by both reviewers every time.
• Reviews are quick. Reviewers think at low and stop after 180 s. With high thinking and no cap they caught the same bugs, but the median clean review took 160 s and the slowest 23 minutes.
• Clean code still draws findings, so Claude has to check each one. Most are minor (36 of 50), but a quarter of clean reviews included a major claim.
• Many of those findings are fair, such as a test that can't fail, an entry point no test covers, or a cache key wider than the change assumed.
• The benchmark already fixed one cause of false alarms. In the first run, reviewers read new files labelled "(untracked)" as left out of the change. Deputy now shows them as git diff does, and those findings went from 6 to 0.
• Agreement is weak evidence on clean code. Both reviewers flagged the same spot on 20% of clean diffs, so a finding isn't real just because two reviewers report it.
• Every finding on a clean diff is listed in the results: default settings, high thinking, no cap, first run.

What it does

Claude (for example Opus 5.5) stays in charge and writes the code. Deputy adds four tools that run a much cheaper model as a helper inside your project:

Tool	What it does
review	Independent reviewers hunt for bugs in a risky or large change before Claude calls it done
consult	Second opinions from 2–4 angles at once, before Claude commits to a plan
investigate	Reads lots of code and returns file:line leads
delegate	Hands a coding task to the cheap model (off by default; on request, or in savings mode)
stats	Shows roughly how many tokens and dollars Deputy saved: this session, today, this week, all time

Claude treats everything Deputy says as an unverified hint: it checks each finding against the code and keeps only what holds up.

See what you saved

Every result ends with a one-line estimate, for example from a real investigate run:

Savings (rough): this work would cost about $0.0082 on Claude Opus 5.5; Deputy did it for $0.0001. Saved about $0.0074 and 2.0k Claude Opus 5.5 tokens.

Ask Claude for stats, or run deputy stats, for totals. Reviews and second opinions are listed separately as extra checks, so they never inflate the savings. The estimate assumes your lead model would use about as many tokens, so treat it as an order of magnitude. Only numbers are stored, never code or prompts. Set your lead model with DEPUTY_LEAD_MODEL (opus, sonnet, haiku, fable or custom).

When to use Deputy, and when not to

Use it for:

• Risky or large changes, where a missed bug costs more than a minute of waiting: tricky logic, security, data handling, concurrency, refactors that touch several files.
• A second opinion before Claude commits to a design, without spending Claude tokens on it.
• Wide reading in a big codebase, such as "where is this setting read?".
• Well-specified mechanical work (boilerplate, renames, test scaffolding) in lean, balanced, partner or savings mode, when you want to save Claude tokens on larger jobs. On small tasks it saved nothing in the benchmark.

Skip it for:

• Routine changes whose tests pass, and trivial edits like typos and one-liners. A review makes you wait and adds little there.
• Code you may not share with a third party. Deputy sends code to DeepInfra, or to whatever endpoint you set in DEPUTY_BASE_URL (a local model works too).
• Final answers. Findings come from a weaker model and include false positives, which is why Claude verifies every one.
• Architecture and security-sensitive decisions in savings mode. Keep those with Claude.

Quick start

1. Get a DeepInfra API key (DeepInfra hosts DeepSeek's models in the US)

1. Sign up at deepinfra.com with GitHub, Google or email.
2. In Dashboard → Billing, add a card and $5–10 of credit, then set a spending limit equal to that amount.
3. In API Keys, create a key (it starts with sk-di-) and copy it. Keep it private.

You don't need a plan, a model purchase or a dedicated GPU. Pay-as-you-go credit is enough.

2. Install uv and the plugin

claude plugin marketplace add alimohammadiamirhossein/deputy
claude plugin install deputy@deputy

Claude Code asks for your DeepInfra key and stores it in your system's secure storage.

3. Restart Claude Code and try it

/deputy:review
/deputy:consult I plan to cache sessions in Redis. What am I missing?
/deputy:investigate how does a request reach the database?

Or just work normally: Claude calls Deputy on its own when it pays off, and skips reviews on routine changes whose tests pass.

Cursor and Codex

One command adds Deputy to the editor's MCP config (with a backup of the old file) and installs the playbook as a Cursor rule or Codex skill:

uvx --from git+https://github.com/alimohammadiamirhossein/deputy deputy setup cursor --project .
uvx --from git+https://github.com/alimohammadiamirhossein/deputy deputy setup codex

Set DEEPINFRA_API_KEY in your environment first, then restart the editor. Add --dry-run to see the changes without writing them. Ready-made files are in integrations/.

Grok instead of DeepSeek (experimental, not tested yet): Grok support follows xAI's documentation but has not been run with a real xAI key. If you try it, please tell us what worked and what didn't, even if it was just a one-word "works". Add --provider xai and set XAI_API_KEY (from console.x.ai) instead. Deputy then uses grok-4.3; set DEPUTY_MODEL to grok-build-0.1 (xAI's coding model) or grok-4.7 (its strongest). Grok costs about 10 times more per token than DeepSeek V4.1 Flash.

uvx --from git+https://github.com/alimohammadiamirhossein/deputy deputy setup cursor --project . --provider xai

Other MCP clients (Cline, Windsurf, Claude Desktop)

Deputy is on PyPI as deputy-cc. Add a stdio MCP server that runs:

uvx deputy-cc serve

with DEEPINFRA_API_KEY set in its environment. Check your setup with uvx deputy-cc doctor.

Recommended model: DeepSeek V4.1 Flash, thinking high

The default is deepseek-ai/DeepSeek-V4.1-Flash-1017 on DeepInfra with thinking set to high. It's the best price-to-quality option we found for this job:

• Cheap: $0.098 per 1M input tokens and $0.294 per 1M output tokens (October 2026), cheaper than DeepSeek's own API.
• Capable: about 1M tokens of context and reliable tool calling.
• Thinking actually on: DeepInfra leaves its thinking off unless asked. Deputy always asks (high), which we confirmed by measuring the reasoning tokens.
• Quick reviews: review and consult think at low by default. In the review benchmark it caught the same planted bugs as high in less than half the time. Set DEPUTY_LENS_THINKING=high for deeper, slower reviews.
• Private: DeepInfra says it doesn't train on your data. Your code isn't sent to DeepSeek's own servers.

Modes

Mode	Who writes the code	Pick it when
accuracy (default)	Claude, always	You want the best result
lean	Claude for small and tricky changes; Deputy reads code and takes larger well-specified work	You want to spend as few Claude tokens as possible with little loss of accuracy
balanced	Claude; Deputy takes boilerplate	You want some savings without risk
partner	Deputy does all low-level work; Claude checks every diff	You want Deputy doing the legwork and pitching ideas, with Claude verifying
savings	Deputy, Claude verifies	You want to save the most tokens

Change it in the plugin settings (/plugin) or with DEPUTY_MODE.

Learn more

• SKILL.md: the playbook Claude follows (consult → decide → implement → review → check each finding). Also works without the plugin.
• docs/REFERENCE.md: all settings, costs, safety, terms of service, the research behind the design, troubleshooting and the roadmap.
• CHANGELOG.md: what changed in each release.

Deputy stays within Anthropic's rules: Claude Code runs unmodified, and Deputy never touches your Claude login. It's an ordinary MCP tool using your own DeepInfra key.

Support the project

If Deputy caught a bug for you, give it a ⭐ and share it with someone who uses Claude Code. Issues and pull requests are welcome; the good first issues are a good place to start.

---

MIT © 2026 Amirhossein Alimohammadi · Not affiliated with Anthropic, DeepSeek or DeepInfra.

Keywords: Claude Code plugin, MCP server, Model Context Protocol, AI code review, second opinion, DeepSeek V4.1 Flash, DeepInfra, subagent, Opus, agentic coding, reduce Claude Code cost.