I beat all the popular harnesses on FrontierHarness Eval

Maki v0.5.5, an AI coding agent built by developer tontinton, topped the FrontierHarness Eval benchmark, according to a report published September 19, 2026. Maki's index tool uses tree-sitter to parse source files into line-numbered skeletons, adding 59 tokens per turn while saving 224 tokens per turn on read calls, a net saving of 165 tokens per turn. The agent also offers a code_execution tool via pydantic's monty interpreter, a task tool for delegating to weak, medium, or strong models, and optional rtk integration that cuts bash output tokens by roughly 50%.

An AI coding agent optimized for minimal use of context tokens, while providing a great user experience. FrontierHarness Eval https://frontierharness.org/ benchmark result, click here to download report.zip https://github.com/tontinton/maki/releases/download/v0.5.5/frontierharness-report-2026-09-19.zip - index tool - uses tree-sitter https://tree-sitter.github.io/tree-sitter to parse supported programming languages to produce a high level skeleton of a file, with exact start-end lines of each item e.g. a function's implementation is in lines 150-165 . Encouraged to be used before reads. For my usage it adds 59 tok/turn but saves 224 tok/turn on read calls, saving 165 tok/turn. - code execution tool - uses monty https://github.com/pydantic/monty to run an interpreter that has all other tools available as async functions. Maki uses it to filter / summarize / transform / pipe data to other tools as input, without it ever reaching and polluting the context window. Sandbox limited by time & memory. - task tool - when delegating work to subagents, the AI chooses whether to run weak / medium / strong model of used provider. Think haiku / sonnet / opus. - System prompt, tool descriptions, and tool examples are all concise, I've made sure not to bloat your context. - Uses rtk https://github.com/rtk-ai/rtk if you have it installed, disable with maki.setup { agent = { rtk = false } } in your init.lua . Saves ~50% of bash output tokens. Remember bash is just 12% of total token usage, so 6% is nice, but saving on reads 65% of total by using index gave me more benefit. I think I'll do bash output filtering like this myself in a future release. - SUPER fast startup, 60 FPS, and light on memory. Not running any JavaScript, using ratatui https://ratatui.rs for TUI. Even the splash screen animation uses SIMD. - Extend with neovim like Lua plugins - Builtin plugins https://github.com/tontinton/maki/tree/main/plugins , User made plugins showcase https://github.com/tontinton/maki/discussions/452 , Lua API reference https://maki.sh/docs/lua-api/ . - Philosophy of not hiding anything - while other coding agents hide information as models improve e.g. not showing number of lines read , maki leaves you in control. - UI fits everything well on my small screen laptop. - Full visibility of subagents - each subagent gets their own "chat window" you can easily navigate between using /tasks Ctrl-X . - Sensible permission system - when the agent runs git diff && rm -rf / , what do you think will happen in your current coding agent? It will treat it as git . Maki uses tree-sitter to parse the bash command and figure out the permissions requested are git and rm . Disable using --yolo . - SSRF protection on webfetch calls. - A memory tool to keep long term context, just tell maki to remember something sometimes it uses it automatically . Managed via /memory view / edit / delete memories . - Fuzzy search with Ctrl-F. - /btw to run a command with the chat history without interfering with the current session. - Rewind on Escape-Escape no code rewind yet, only chat history . - Attach images in prompts. - 26 of the most popular themes. - Resume sessions. - Skills & MCPs. - Opt-in OpenTelemetry https://maki.sh/docs/telemetry/ export, same format as Claude Code's. - Plan mode. - Run bash commands using , or if you want maki to not know about it. - /cd to change dir. - Use --print --output-format stream-json to run UI-less. Output is compatible with Claude Code, so you can easily replace your existing solutions. - Anthropic - ANTHROPIC API KEY only using OAuth is against TOS . Bedrock supported via CLAUDE CODE USE BEDROCK=1 . - OpenAI - OPENAI API KEY and OAuth via maki auth login openai . - xAI - XAI API KEY and OAuth via maki auth login xai . - Google - GEMINI API KEY . - Copilot - GH COPILOT TOKEN or an existing GitHub Copilot sign-in at ~/.config/github-copilot/ . - Ollama - OLLAMA HOST for local e.g. http://localhost:11434 , or OLLAMA API KEY for cloud. - llama.cpp - LLAMA CPP HOST e.g. http://localhost:8080 , optionally LLAMA CPP API KEY . - Mistral - MISTRAL API KEY . - Z.AI - ZHIPU API KEY . - DeepSeek - DEEPSEEK API KEY . - OpenRouter - OPENROUTER API KEY . - Requesty - REQUESTY API KEY . Set REQUESTY BASE URL=https://router.eu.requesty.ai/v1 for the EU region. - Synthetic - SYNTHETIC API KEY . - Regolo - REGOLO API KEY . EU-hosted open-weight models. - TensorX - TENSORX API KEY . - OpenCode Zen - OPENCODE API KEY , or the free public key for zero-cost models. Models from the models.dev catalog. - OpenCode Go - OPENCODE API KEY . Models from the models.dev catalog. - Aperture - APERTURE HOST e.g. https://your-host.tailnet.ts.net . No API key needed, Tailscale handles auth. Dynamic providers - drop an executable script into ~/.config/maki/providers/ to add custom providers or proxies. See docs https://maki.sh/docs/providers/ dynamic-providers for details. Providers as Lua plugins WIP Download and read the script first don't blindly trust shell scripts . curl -fsSL https://maki.sh/install.sh -o install.sh cat install.sh Then run. chmod +x install.sh && sh install.sh One-liner: curl -fsSL https://maki.sh/install.sh | sh Installs to ~/.local/bin . Override with MAKI INSTALL DIR . Download and read the script first don't blindly trust remote scripts . irm https://maki.sh/install.ps1 -OutFile install.ps1 Get-Content install.ps1 Then run. .\install.ps1 One-liner: irm https://maki.sh/install.ps1 | iex curl -fsSL https://maki.sh/install.sh | sh Both install to %LOCALAPPDATA%\maki and add it to your user PATH. Override with MAKI INSTALL DIR / $env:MAKI INSTALL DIR . cargo install --locked --git https://github.com/tontinton/maki.git maki nix run github:tontinton/maki Or download a pre-built binary from GitHub Releases https://github.com/tontinton/maki/releases/latest . Run maki acp or configure your ACP supporting editor to use maki, e.g. in Zed https://zed.dev/ 's settings.json : "agent servers": { "Maki": { "default config options": { "model": "deepseek/deepseek-flash" }, "type": "custom", "command": "maki", "args": "acp" , "env": {} } } More info at the official docs https://maki.sh/docs . tontinton/makiconf https://github.com/tontinton/makiconf - includes a semble https://github.com/MinishLab/semble tool Lua code for semantic code search, and an ast-grep https://ast-grep.github.io MCP server for AST-based search and replace. DISCLAIMER: 90% of code in maki was written by maki, guided by humans. Some parts of the code are not as good as what I would've made in the artisanal hand-made style. But it's also not slop / vibe coded, and can easily be refactored if needed nowadays.