CrabTalk: An 8MB Open Source Agent Daemon Written in Rust
CrabTalk is a lightweight, open-source agent daemon that runs agents, dispatches tools, and connects to MCP servers — all in an 8MB binary written in Rust. Released v0.0.15 on March 28, the project has gained 378 stars on GitHub since its creation on March 11, 2026.
The daemon ships with built-in tools for shell execution, task delegation, and memory management, plus native MCP server integration. Skills are defined as Markdown prompt files with YAML frontmatter, organized across three tiers: builtin, user, and project. Heavier capabilities live outside the daemon as independent components that auto-discover via PATH.
What makes CrabTalk distinctive is its transparency philosophy: the daemon streams every agent event to your client — text deltas, tool calls, thinking steps, all of it. There is nothing hidden. The architecture separates the lightweight core from extensible components, keeping the binary small while allowing unlimited capability expansion.
Installation is simple: `cargo install crabtalk` or a one-line shell script. The project is dual-licensed under MIT and Apache-2.0, and ships with comprehensive documentation including architecture docs and RFCs.
For teams building agent-powered workflows who want full visibility into what their agents are doing, CrabTalk offers a minimal-footprint alternative to heavier frameworks.
GitHub: [github.com/crabtalk/crabtalk](https://github.com/crabtalk/crabtalk) | Website: [crabtalk.ai](https://crabtalk.ai)
← Back to all articles
The daemon ships with built-in tools for shell execution, task delegation, and memory management, plus native MCP server integration. Skills are defined as Markdown prompt files with YAML frontmatter, organized across three tiers: builtin, user, and project. Heavier capabilities live outside the daemon as independent components that auto-discover via PATH.
What makes CrabTalk distinctive is its transparency philosophy: the daemon streams every agent event to your client — text deltas, tool calls, thinking steps, all of it. There is nothing hidden. The architecture separates the lightweight core from extensible components, keeping the binary small while allowing unlimited capability expansion.
Installation is simple: `cargo install crabtalk` or a one-line shell script. The project is dual-licensed under MIT and Apache-2.0, and ships with comprehensive documentation including architecture docs and RFCs.
For teams building agent-powered workflows who want full visibility into what their agents are doing, CrabTalk offers a minimal-footprint alternative to heavier frameworks.
GitHub: [github.com/crabtalk/crabtalk](https://github.com/crabtalk/crabtalk) | Website: [crabtalk.ai](https://crabtalk.ai)
Comments