CodeGraff (graff) is an AI coding agent for the terminal, shipped as a single binary of about 3 MB with no runtime dependencies, written in Zig for macOS, Linux, and Windows. You describe a task in plain English and graff reads and edits files, runs commands, uses browser tools, and hands parts of the work to sub-agents.
Sub-agents run in parallel, each with its own context, and a workflow runs phases of parallel children in sequence, passing each phase's results to the next. Context stays small: stable setup is reused, large tool outputs become handles the model pages through, and a compact command shortens the transcript.
Graff runs interactively, as a one-shot prompt, as an HTTP session server, as an MCP server that other clients hand small tasks to, and as an Agent Client Protocol (ACP) agent inside editors such as Zed. Generated Python and TypeScript SDKs drive it from code. Harness, a separate native desktop app, bundles graff alongside other installed coding agents with chat, file review, a browser, and terminals in one window.
Recorded evaluations on 12 live pull-request tasks (three runs each, pass at two of three) report these results with the same grok-4.6 model:
| Harness | Tasks passed | List price | Mean wall time |
|---|---|---|---|
| graff | 12/12 | $21.48 | 264s |
| Pi | 12/12 | $18.47 | 334s |
| OpenCode | 12/12 | $25.71 | 309s |
| grok | 11/12 | $33.69 | 362s |
| exo | 9/12 | $16.01 | 281s |
Measured footprint is a cold start of about 1.8 ms, about 12 MB peak memory for a full agentic turn, and about 0.4 MB per additional parallel sub-agent. A separate evaluation suite of 45 tasks also runs the 30 FrontierHarness tasks from Terminal-Bench 2.1 and DeepSWE under a documented, non-identical protocol.
Features
- Single binary: about 3 MB with zero runtime dependencies, for macOS, Linux, and Windows
- Sub-agents and workflows: parallel sub-agents with separate contexts, and workflows that chain phases of parallel children
- Context handles: large tool outputs become small handles the model pages through instead of filling the transcript
- MCP client: adds local or remote MCP servers by name, URL, or command and checks each one works before saving it
- Embeddable: runs as an HTTP session server, an MCP server, or an ACP agent for editors such as Zed, with generated Python and TypeScript SDKs
- Permission modes: asks before writes by default, with a read-only plan mode, a strict mode, and a skip-all-prompts mode for CI
- Many providers: Anthropic, OpenAI, DeepSeek, xAI, Z.AI, Kimi, MiniMax, Xiaomi, Mistral, OpenRouter, Groq, and Cerebras, plus ChatGPT plan sign-in
- Claude Code import: imports a project's Claude Code conversations so they can be resumed in graff
- Harness desktop app: a native desktop app that runs graff or another coding agent with chat, review, files, a browser, and terminals
- Evaluation suites: live pull-request tasks, 45 in-house tasks, and a FrontierHarness runner with recorded results
