Maki is an open-source coding agent for the terminal: it reads codebases via a tree-sitter-based index tool, edits files, runs commands in a sandbox, and connects to a wide range of model providers to turn any language model into a working AI harness in the terminal. Maki is developed by Tony Solomonik (GitHub: tontinton) as an MIT-licensed open-source project. As of 20 September 2026.
What is Maki?
According to its official GitHub repository, Maki comes from Tony Solomonik and is released under the MIT licence; the repository was created in February 2026. Maki is installed via curl -fsSL https://maki.sh/install.sh | sh (Linux/macOS) or a PowerShell script on Windows, or alternatively via cargo install --locked --git https://github.com/tontinton/maki.git maki or nix run github:tontinton/maki. It is started by running maki inside a project directory. Its interface is a lightweight terminal UI built on ratatui, with no JavaScript runtime, and according to the README a high refresh rate and low memory footprint. Via the Agent Client Protocol (ACP, maki acp), Maki can also be integrated into ACP-capable editors such as Zed.
Maki as an AI harness
Per its own description, Maki aims to get by on as few context tokens as possible: the index tool uses tree-sitter to compress files into a skeleton of imports, type definitions and function signatures with exact line ranges, which according to the README makes reads considerably cheaper than full file reads. The code_execution tool runs further tools inside a time- and memory-bounded sandbox via the monty interpreter, without writing intermediate results into the visible context. Permissions are not granted by tool name alone: tree-sitter parses bash commands and, per the README, distinguishes git * from rm * permissions in a command like git diff && rm -rf /; --yolo disables these prompts entirely. A memory tool, managed via /memory, retains durable context across sessions. Via the task tool, Maki delegates subtasks to subagents with a selectable model strength (weak, medium or strong) and its own chat window reachable via /tasks. Extensibility comes through skills, MCP servers and neovim-style Lua plugins, configured in init.lua.
Models and model access
Maki is deliberately model-agnostic: according to the README, it supports Anthropic Claude models via ANTHROPIC_API_KEY, OpenAI GPT models via OPENAI_API_KEY or OAuth, plus xAI, Google Gemini, GitHub Copilot, Mistral, DeepSeek, OpenRouter and local models via Ollama and llama.cpp; for Anthropic, the README explicitly states only API-key access is intended, since OAuth would violate Anthropic’s terms of service. Via environment variables such as OLLAMA_HOST or LLAMA_CPP_HOST, or a custom script in ~/.config/maki/providers/, Maki can also be pointed at a self-hosted, OpenAI-compatible endpoint – which makes it possible to run Maki through innFactory’s AI Gateway, a central, OpenAI-compatible proxy for agent harnesses with budgets, cost centres and authentication via Microsoft Entra ID.
Enterprise use: data protection and governance
Maki is locally installed open-source software with no central operator; the documentation does not describe any servers or data processing run by the project itself. Prompts and source code go directly to the model provider the user chooses – Anthropic, OpenAI, Google and the other supported providers each have their own privacy and retention terms that should be checked in advance. Per the README, telemetry is disabled by default and can optionally be enabled as an OpenTelemetry export in the same format used by Claude Code. Central enterprise features such as SSO, an admin console or a data processing agreement are not documented for this community project. Important in practice: source code and prompts are sent to whichever model provider is configured – this data flow belongs in an organisation’s records of processing activities in advance.
Orchestrating Maki with herdr
According to the documentation of the terminal orchestrator herdr, it detects Maki sessions via a screen manifest, without accessing any state or session information from Maki itself. This allows a basic status view of Maki instances alongside other coding agents in one shared terminal interface, but not the deeper integration herdr documents for other supported agents.
Where it fits: developer harness and enterprise harness
Maki is a harness for developers who want to work in their own terminal as context-efficiently and provider-independently as possible. For business users outside development, CompanyGPT is the matching harness: tools via MCP servers and n8n workflows, context via companyRAG, roles from Entra ID or Keycloak, and audit logs for business departments. The AI Gateway forms the shared cost and governance layer for both worlds. For an overview of other coding agents and orchestrators, see our overview of all AI harnesses. We help organisations adopt open-source harnesses such as Maki securely too, with proper governance and centralised cost control through an AI Gateway.