Cheap Agent
Let a frontier model plan while a low-cost model implements code changes and proves them with tests.
- Source repo
- muaaz9329/cheap-ag
- Stars
- ★ 339
- Last updated
- 5d ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 57/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeOpenAI API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A Claude Code or Codex developer who wants a low-cost model to implement routine feature work.
- Not a fit if
- Teams unwilling to send project code to their chosen model provider
- Developers who want the frontier model to write every line
- Website
- cheapag.com ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-16latest release v1.4.3 (2026-11-16)+1.3 ★/day over 7 days
- Source review
- 57/100 · Major gaps
What does this agent do, and when should you use it?
Cheap Agent is a CLI, MCP server, and plugin for Claude Code and Codex, running on Node.js 20+. The frontier model plans a task, while a low-cost model writes code and tests in a Git worktree copy of the project. Cheap Agent runs the tests itself; changes that pass its checks land back in the project uncommitted for review. Providers include OpenRouter, DeepSeek, OpenCode Go, Command Code, OpenCode Zen, and custom OpenAI-compatible endpoints. Before using it in a project, you run the relevant allow command to consent to sending that project's code to the selected model.
Run npx cheap-agent to choose a provider and models and enter an API key; a custom endpoint or local model can also be configured. Then run /cheapag:allow in Claude Code or $cheapag:allow in Codex to authorize the current project. After the frontier model assigns a task, Cheap Agent has the low-cost model implement the change and write a test that fails before the fix and passes after it in a Git worktree. It runs the tests itself. Changes passing its checks land back in the project uncommitted; an unverified reproduction lands with a risk note. Use /cheapag:status for recent jobs and cost, /cheapag:model to choose primary and fallback models, and /cheapag:provider to switch among providers already set up.
- A Claude Code or Codex developer who wants a low-cost model to implement routine feature work.
- A team that wants Opus to plan substantial changes while reducing how much Opus writes.
- A maintainer who needs to try fixes in an isolated Git worktree and review the resulting uncommitted diff.
- A developer who wants tests run before accepting changes implemented by a low-cost model.
- A developer with a local model or custom OpenAI-compatible endpoint who can configure its model details.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npx cheap-agentnpx cheap-agent watch
- Runtime
- Node.js 20+
- Install notes cover
- Windows
How do you install or deploy this agent?
Requires Node.js 20+, Claude Code or Codex, and an API key for the chosen provider (a custom endpoint may not need a key). Run in a terminal:
npx cheap-agentChoose a provider and models in the prompts, then enter your API key. If you use Codex only, decline the offer to install into Claude Code, then add the Codex plugin:
codex plugin marketplace add muaaz9329/cheap-ag
codex plugin add cheapag@cheap-agentOn Windows, Git for Windows (Git Bash) is also required.
How do you use this agent?
In a Claude Code project, authorize it once:
/cheapag:allowIn a Codex project, start Codex, trust the 3 cheapag hooks when prompted, and run:
$cheapag:allowThen request code changes as usual. Use the corresponding /cheapag: or $cheapag: commands for status, model selection, and provider switching. To watch progress from another terminal, run:
npx cheap-agent watchWhat are this agent's strengths and limitations?
- Supports both Claude Code and Codex, with MCP server and Codex plugin integration paths.
- The low-cost model works in a separate Git worktree; changes that pass checks land uncommitted for review.
- Supports several hosted providers, custom OpenAI-compatible endpoints, and local models, with provider switching among configured accounts.
- The README reports that both configurations solved all 7 everyday SWE-bench Verified fixes, while the Cheap Agent configuration used 59% fewer Opus tokens.
- Requires Node.js 20+, Claude Code or Codex, and a configured model provider; hosted models generally require an API key and usage charges.
- After explicit project consent, project code is sent to the chosen model provider, which may not suit projects that prohibit this transfer.
- The README reports 6 of 9 hardest benchmark bugs solved with Cheap Agent versus 7 of 9 with plain Opus; tiny tasks may offer no savings.
- Windows also requires Git for Windows, and the first job may take about two minutes while the OpenCode engine sets up.
How does this agent compare with similar options?
The README compares it with Claude Code + Opus alone: in the described SWE-bench Verified run, everyday fixes used 59% fewer Opus tokens, while the hardest set had 6/9 solved versus 7/9 by plain Opus. For a step the low-cost model cannot finish, Claude Code can ask, hand over to a Sonnet helper, or continue itself depending on settings; Codex asks in chat or continues itself.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Cheap Agent This agent | 57 · Major gaps | CLIFree + model costs | ★ 339 | 5d ago | TypeScript | Codex · Claude Code |
| Agent Deck | 74 · Some gaps | CLIFree + model costs | ★ 1.3k | 1d ago | Go | Codex · Claude Code |
| ECC Engineering Harness | 63 · Some gaps | Agent plugin / skillFreemium | ★ 308k | 1d ago | JavaScript | Codex · Claude Code |
| Puppetmaster | 73 · Some gaps | CLIFree + model costs | ★ 479 | today | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Trust: Least privilege — the skill restricts allowed-tools to npx cheap-agent subcommands and MCP tools run over stdio, with no excessive permissions observed, but server.ts is truncated so the full tool surface cannot be confirmed, hence 2. User confirmation — /cheapag:allow explicitly records project-level consent, takeover defaults to ask, and the model is explicitly forbidden from self-approving consent, which is good; however the hook edit guard intercepts Claude edits, a fairly strong intervention that is only partially explained. Data flow transparency — the README states code is sent to the chosen provider only after allow and shows the model, but there is no field-by-field data flow inventory. Sensitive data handling — API keys use hidden prompts and local credential storage, redact() wraps all tool output, and custom endpoints reference only env var names or files without reading secrets, which is good; but the credential storage implementation is not in the sample. Dependency security — dependencies include @opencode-ai/sdk, opencode-ai pinned at 1.19.0 and the MCP SDK, with no lockfile or vulnerability audit evidence, hence 1. External effects — no commit/push is explicit, changes land uncommitted, and the worktree is isolated, which is good. Rollback — a clean command, retention cleanup, worktree isolation and uncommitted landing make discarding easy, but there is no explicit mechanism to undo landed changes. Source attribution — MIT and author attribution are present, but publisher identity is unverified and SWE-bench data points only to a personal blog, so attribution evidence is weak.
Reliability: README, SKILL.md and package.json version 1.4.1 match the [email protected] pinned in the skill, and the command table broadly matches the index.ts subcommand dispatch, so consistency is good; but server.ts is truncated, so all tools and instructions cannot be checked. Dependency availability — dependencies are clearly declared and engines require node>=20.11, but the sample has no package-lock, the opencode-ai engine downloads on first run, and Windows is known to leave residual processes, so availability carries external dependency risk, hence 1. Failure messages — instructions define structured result codes such as OK/PARTIAL/FAIL/BLOCKED/CONFLICT/TIMEOUT/BUDGET/PROVIDER/CONSENT with handling guidance, so failure messaging is fairly complete.
Adaptability: Audience and scenarios — targets both Claude Code and Codex hosts, covers students, budget-sensitive users and enterprise use, and supports local models and custom endpoints, so scenario coverage is broad. Capability boundaries — explicitly states no benefit on small tasks, slightly lower success on the hardest bugs, Windows limitations and weak tool calling in small local models, so boundaries are clearly stated. Trigger precision — skills set disable-model-invocation and require explicit user invocation; MCP instructions explicitly require code changes to go through delegate, so trigger rules are specific. Environment fit — supports macOS/Linux/Windows, requires Git Bash on Windows and lists known issues, so environment fit is well documented.
Convention: Information architecture — the README is clearly structured with install, commands, settings, FAQ and update/removal sections, so information architecture is good. Install notes — provides an agent auto-install script, step-by-step commands for Claude Code and Codex, and dependency requirements, so install notes are detailed. Naming stability — package name cheap-agent, command cheapag and plugin cheapag are consistent, but the README uninstall commands show both cheapag@cheapag and cheapag@cheap-agent, a minor inconsistency. Examples and FAQ — includes before/after comparison tables, FAQ and demo images, so examples are sufficient. Known limitations — explicitly lists hardest-bug performance, no benefit on small tasks and Windows issues, so limitations are well disclosed. License — full MIT text is present and matches the package.json declaration. Versioning and changelog — a version number and Releases page pointer exist, but the sample has no CHANGELOG file, hence 1. Maintenance responsibility — author attribution and Reddit/blog channels exist, but the publisher is unverified, so evidence for maintenance ownership is limited.
Effectiveness: Output usability — output is returned as structured result blocks (OK/PARTIAL/FAIL etc.) and changes land as an uncommitted diff for easy review, so output usability is good. Marginal value — can significantly reduce frontier-model tokens on real feature and fix scenarios, and the README gives before/after comparisons, so the value proposition is concrete; but small tasks see no benefit, so value is conditional. Cost benefit — claims 63% token reduction and half-cent cost, but the data comes from the author's single run and personal blog with no independent reproduction, so cost-benefit evidence is assertion, hence 1.
Verifiability: Claim traceability — SWE-bench and token figures point to the author's blog and charts, with no reproducible raw logs or harness output provided, so traceability is weak. Cross-source corroboration — README numbers match the assets charts, but all come from the same author source with no independent third-party corroboration. Fact/inference separation — the README partly admits limitations (hardest bug 6/9 vs 7/9), showing some honesty, but marketing claims (50x, 63%) are mixed with measured details, so separation is incomplete.
- Publisher identity is unverified; verify npm package and GitHub ownership before installing.
- SWE-bench and token-savings figures come from the author's blog and single runs, not independently reproduced; treat as unverified performance claims.
- The opencode-ai engine downloads on first run, and on Windows a crash may leave an opencode.exe process that must be cleaned up manually.
- The hook edit guard blocks large Claude edits and forces delegation; understand its behavior before enabling.
- server.ts is truncated in the sample, so the full MCP tool surface and permission boundary cannot be fully reviewed.
- README uninstall commands show both cheapag@cheapag and cheapag@cheap-agent, which may confuse users.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v1.4.3 1.4.3
Release notes excerpt
Ollama is a provider. Pick Ollama at login (number 4), then one of two endpoints: - This machine: no key. Login lists the models you've pulled and reads each one's context window from Ollama. Nothing leaves your machine, so even /cheapag:allow --zdr-only allows it. - Ollama Cloud: export OLLAMA_API_KEY where Claude Code starts, then log in and pick a preset. Like the other providers, jobs switch to another cloud model when one hits a limit, and move up to a stronger one when stuck. If a cloud job can't find its model, pick one from your account with /cheapag:model. …View on GitHub ↗ -
v1.4.1 1.4.1
Release notes excerpt
Fewer failed jobs, and you decide before Claude spends tokens. When the cheap worker can't finish, Claude now asks you first: continue with Claude, retry the worker on another model, or stop. Before, it handed the step to a Sonnet subagent on its own, often costing several hundred thousand tokens. The old behaviour is still available with takeover: "sonnet". …View on GitHub ↗ -
v1.3.2 1.3.2
Release notes excerpt
Updating is one command again. npx cheap-agent@latest on a machine that's already set up now just updates: no provider or key questions, the Claude Code server entry moves to the new version (before, it stayed pinned to the old one, so only the plugin updated), and the plugin updates. It prints what it did and tells you to restart Claude Code. npx cheap-agent login still runs the full setup when you want to change provider or models. If you're on 1.2.7 or 1.3.1: the update notice already says to run npx cheap-agent@latest, and that command now does the right thing. …View on GitHub ↗
FAQ
Where does my code go?
/cheapag:allow or $cheapag:allow in that project; the consent step shows which models are covered. The API key stays in a local credential store rather than chat.Do I need a paid model account?
What happens if the low-cost model gets stuck?
takeover setting can prompt you, route the step to a Sonnet helper, or let Claude finish it. Codex asks in chat or finishes it itself.Will it commit or push changes?
Can I use a local model?
Is Cheap Agent open source?
Is Cheap Agent safe to use?
Is Cheap Agent still maintained?
Who makes Cheap Agent?
What is Cheap Agent's official website?
Thanks — we use this to improve the page.