An open-source coding agent developed by Thomas Wood. It natively supports the Agent Client Protocol (ACP), so the same agent runs standalone in a terminal or inside an ACP-compatible editor such as Zed. Its design philosophy is “zero framework” — instead of inventing a new agent framework, it is assembled from existing parts such as the OpenAI SDK and FastMCP. It uses a two-tier structure: crow-cli handles the thinking, and crow-mcp actually touches files and runs commands. Seven built-in tools, including web search, are available from the start. It is released under the MIT license and was created as an alternative to subscription-based agent services.
Key Features
- ACP-native, moving freely between CLI and editor: Because it follows the Agent Client Protocol rather than a proprietary protocol, any ACP-compatible client can connect to it directly. The developers point to a fork of Zed as their recommended client
- Two-tier architecture separating thinking from execution: The ACP agent itself,
crow-cli, acts as the “brain” running a ReAct loop, while the MCP tool servercrow-mcpacts as the “hands” that touch files, the shell, and the web. This makes it easier to swap out the tool layer or use it from other MCP clients - Seven built-in tools: File operations (read, write, and edit with fuzzy string matching), terminal command execution, web search via SearXNG, fetching web pages as Markdown, webcam capture, reading image files, and memory access are all available out of the box
- Web search without a search API key: Searches are sent to a self-hosted SearXNG metasearch engine (launched via Docker), so there is no need to sign up separately for a dedicated search API key
- Memory that spans sessions: Conversation and tool-execution history is stored locally, and past sessions can be searched and recalled. Agents referencing each other’s history is also part of the intended usage
- Extensible through skills and subagents: Placing a
SKILL.mdfile under~/.agents/skillsmakes that procedure loadable at runtime, and the agent picks the ones that fit the task. A mechanism for delegating work to another agent is also provided - Bring your own LLM: OpenAI-compatible endpoints (OpenAI, OpenRouter, custom endpoints, and so on) are specified in the configuration file
~/.agents/crow/config.yaml. The choice of model stays with the user
Pricing
| Plan | Price | What’s included |
|---|---|---|
| The software itself (MIT license) | $0 | All features. Install from GitHub and use without limits |
| LLM usage | Pay-as-you-go per provider | API charges from the provider you connect to (OpenAI, OpenRouter, etc.) apply separately |
crow-cli itself is free and open-source software, and the official site lists no paid plans. Actual cost depends entirely on which model you send how much work to, so switching to a cheaper model keeps expenses down, while heavy use of high-end models incurs proportional API charges.
Pricing is current as of August 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- MIT-licensed open source with no monthly subscription required; cost is limited to pay-as-you-go API usage
- Being ACP-compliant, the same agent can be reused in a terminal and in a compatible editor
- Search, file, shell, and image tools are included from the start, so practical work is possible without hunting for additional MCP servers
- Self-hosting SearXNG removes the need for a separate service’s API key for web search
- The implementation is a thin combination of existing libraries, making it easy to read the code, verify behavior, and modify it
⚠️ Cons
- Setup requires Python and Docker, a higher barrier than tools that work entirely through a GUI
- You must supply your own LLM provider API key and manage usage costs yourself
- It is an actively developed individual project, and the community is smaller than that of major vendors’ agents. Documentation and implementation may diverge in places
- ACP-compatible clients are still limited, so editor integration options are few
- There is almost no substantial Japanese-language material; reading the English documentation and source code is assumed
Comparison with Similar Services
| Criteria | crow-cli | Claude Code | Gemini CLI | Aider |
|---|---|---|---|---|
| Provider | Thomas Wood (individual, OSS) | Anthropic | OSS community | |
| License | MIT (open source) | Proprietary | Open source | Open source |
| Usable models | Any OpenAI-compatible model | Claude series | Gemini series | Choice of major providers |
| Editor integration | ACP-compatible clients (Zed, etc.) | Dedicated extensions and IDE integration | Supported editors and IDE integration | Mainly terminal |
| Pricing model | Free software + pay-as-you-go API | Subscription or pay-as-you-go API | Free tier + pay-as-you-go API | Free software + pay-as-you-go API |
All of these delegate coding work from a terminal, but crow-cli positions itself as “a lightweight implementation that rides on protocol standards (ACP and MCP) without being tied to one vendor’s models.” It suits people who value model choice and readable code throughout.
Who Is It For
- People who want a coding agent but prefer pay-as-you-go API billing over a monthly subscription
- People who want to pick and switch models themselves across OpenAI, OpenRouter, and others
- People using Zed or another ACP-compatible editor who want to run an agent inside it
- Developers who want to read and understand an agent’s internals, or modify them themselves
- People who want to delegate work including web search without adding yet another search API key
Summary
crow-cli is a coding agent built thinly on the standard protocols ACP and MCP, holding to a policy of not inventing a new framework. A structure that separates the thinking crow-cli from the executing crow-mcp, key-free web search via SearXNG, and memory that spans sessions are all available under the MIT license. If you can handle the Python and Docker setup, it is well worth considering as an alternative to subscription-based agents. Start by following the installation steps on GitHub and trying small tasks with an API key you already have.