Maki

Maki is an open-source terminal coding agent written in Rust by tontinton. It reads, searches, edits, and runs code in a repository, and is built to spend as few context tokens as possible. It connects to many model providers and is extended through Lua plugins.

What It Is

Maki is a TUI coding agent: point it at a repo, pick a provider, and it works through the task using built-in tools such as read, bash, edit, and batch. It runs as a native binary with no JavaScript runtime, using ratatui for the interface. It also supports headless use via --print with output the docs describe as compatible with Claude Code, and can be driven from editors such as Zed over the Agent Client Protocol.

How it reduces token use

The project's main focus is context efficiency. Several tools work together:

• index parses supported languages with tree-sitter into a skeleton of imports, types, and function signatures with line ranges, so the model reads only the lines it needs.
• code_execution runs a sandboxed Python interpreter where other tools are async functions, so only printed output enters the context.
• tool_search hides large MCP tool sets behind one search tool and loads definitions on demand.
• task lets the model choose a weak, medium, or strong model for each subagent and returns a summary rather than the transcript.
• Compaction drops images and thinking blocks first, then summarizes old turns.

The author states that in the FrontierHarness Eval benchmark Maki 0.5.5 reached a 70% pass rate at $2.06 per pass, and claims it cuts cost and time by about 2x in benchmarks.

Extending with Lua

Maki has a Neovim-style Lua plugin API. Built-in tools are themselves Lua plugins, and users can add tools, slash commands, keymaps, UI, and whole model providers by placing files in the config directory. Skills, MCP servers over stdio or HTTP, hooks, and Lua packages are also documented.

Providers and controls

Supported providers include Anthropic, OpenAI, xAI, Google, Copilot, Ollama, llama.cpp, Mistral, DeepSeek, OpenRouter, and others, plus any API compatible with OpenAI or Anthropic formats. Bash commands are parsed with tree-sitter so permission rules see every command in a compound line, and --yolo skips permissions. Other documented features include long-term memory, plan mode, parallel sessions, subagent chat windows, fuzzy search, double-Escape rewind, image paste, 26 themes, and opt-in OpenTelemetry export.