maki is unclaimed —
M
maki
the efficient coder·maki.sh
Maki is a terminal coding agent written in Rust with a native TUI. It reads, edits, searches, and runs code in a codebase while using context-reduction techniques (skeleton indexing, a sandboxed code-execution interpreter, subagent model tiering, and automatic context compaction) to cut token usage. It connects to LLM providers such as Anthropic, OpenAI, Google, and others via their HTTP APIs, and supports MCP tool servers, Lua plugins, and headless/editor (ACP) integration.
What it's for
reading, editing, searching, and running code in a codebaseparsing codebases into compact structural skeletons instead of reading full filesrunning sandboxed Python scripts that chain multiple tool calls without polluting contextdelegating research or architecture subtasks to subagents using different model tiersconnecting to external tool servers via MCPrunning headless in scripts or CI via --printdriving the agent from an editor like Zed via ACPextending the agent with Lua plugins (tools, commands, keymaps, UI)
Features 22
- ACP server (maki acp) for editor integration
- async tool exposure allowing the model to gather multiple reads/searches in one script
- automatic context compaction: strips images, thinking blocks, summarizes older turns
- bash command parsing via tree-sitter for permission checks
- code_execution: sandboxed Python interpreter with memory and time limits
- fuzzy search (Ctrl-F) and /btw side queries
- headless mode via --print/-p with Claude Code-compatible output
- index tool: parses 15 languages into structural skeletons (imports, type defs, function signatures with line ranges)
- long-term memory persisting across sessions
- Lua plugin API mirroring Neovim's, for adding tools, commands, keymaps, and UI
- MCP server support over stdio or HTTP
- MCP tool_search: hides large MCP tool sets behind a single lookup tool
- native Rust TUI running at 60 FPS with background-thread syntax highlighting
- parallel sessions with /sessions switching
- per-subagent chat windows navigable with Ctrl-N/P
- per-tool allow/deny permission rules, or --yolo to bypass
- plan mode restricting the agent to read-only actions
- shell command execution with ! and !!
- Skills, 26 UI themes, image pasting
- SSRF protection on webfetch
- status bar showing token count, cost, and model at all times
- task tool: subagents assigned weak/medium/strong model tiers, read-only or full tool access