Harness of harnesses  ·  open source

Every agent.
One cockpit.

Run Cursor, Codex, Claude, Muse Code, and OpenCode side by side, from your desktop, your phone, or Discord. Self-hosted on your machines, and yours to reshape.

Get started View on GitHub
  • MIT licensed
  • Windows + Linux
  • Android + Discord
Cuttle  ·  127.0.0.1:8080

Hosts the agents you already use

  • Cursor
  • Codex
  • Claude Code
  • Muse Code
  • OpenCode
  • Hermes
  • DeepSeek
  • Antigravity
  • Ollama & llama.cpp

Tour

A chat that runs your agents.

Pick an agent per chat or let the router choose. Every turn streams progress, keeps a task list, and asks for decisions as clickable cards.

Visible work

Agents that show their work.

Follow each turn as it happens, then approve the next step without typing a word.

  • Collapsible progress notes on every reply
  • Pinned Tasks lists the agent creates and ticks off
  • One-click action forms for questions, approvals, and restarts
Rate-limit uploads

Rich answers

Answers you can read at a glance.

Agents reply with tables and Vega-Lite charts that render inline, so reports look like reports.

  • Markdown tables, Vega-Lite charts, and code blocks
  • Images and PDFs in, with a vision pre-pass for every harness
  • Shared history across web, desktop, phone, and Discord
Agent costs this week

Workspace

Split, resize, switch spaces.

Tile chats and dashboards any way you like. Every space keeps its own layout, so a build, a review, and a render can each have a home.

  • Split any pane horizontally or vertically
  • Drag dividers to resize
  • Desktop tabs for spaces, each with its own layout
Cuttle  ·  spaces

Model benchmarks

Know which model earns its cost.

Compare providers by cost, pass rate, and speed in one interactive 3D view, next to what your own agents actually spent.

  • Model Benchmarks across every provider
  • Per-agent cost and performance tracking
  • A router that escalates when a cheap run fails
Dashboards  ·  Model Benchmarks

Open source  ·  MIT

Make it yours, one prompt at a time.

Cuttle is built to be changed by the agents it hosts. Open a chat on the Cuttle repo, describe what you want, and watch it reshape itself around your workflow.

CuttleCursor

add a Pomodoro timer to the left rail

  • add a Pomodoro timer to the left rail
  • build a dashboard that charts my agent spend by project
  • make a theme from the colors in this screenshot
  • add a /deploy command that uploads my game to itch.io
  • give my Discord bot a daily standup summary
Settings  ·  Theme

No build step

The UI is plain HTML, CSS, and vanilla JS served by Flask. Refresh and the change is live; Python changes apply from a one-click restart card.

Agents arrive briefed

AGENTS.md, Cursor rules and skills, and .cuttle/docs runbooks tell a fresh agent where things live and what not to break.

Customize without forking

A project's .cuttle/ folder adds its own slash commands, clickable actions, rules, and docs. Personal tweaks stay out of git.

Swap any part

Harnesses, the router's brain, and local or cloud models are all sockets. Bring the agents and models you trust.

Features

Everything around the agents.

The parts a single CLI doesn't give you: routing, fan-out, a LAN mesh, and a home for every project.

Agent router

Picks a harness on fresh chats and escalates when a run fails: Cursor Auto, then Grok, then the Codex fallback chain. Stop is always final.

Sub-agents

Fan a task out into real child chats across harnesses for votes, critiques, or parallel research, then get one synthesized reply.

Cuttle Workers

Every PC on your LAN becomes a worker: file copy, shell recipes, a Blender render farm with work stealing, and self-update.

Projects that brief agents

Each project's .cuttle/ rules, commands, and docs are compiled into every turn by Cuttle Brain, whichever agent runs it.

Desktop, phone, Discord

An Electron Host and LAN Clients, an Android app, and Discord DMs all open the same chats and history.

Mid-turn steering

Send a follow-up while Codex or Muse Code is working and it lands in the live turn instead of waiting in a queue.

Local models

Ollama and llama.cpp paths for cloud-off work, with requests queued so one GPU never gets swamped.

Agent ops CLIs

Agents drive Cuttle through python -m api.* verbs for chats, widgets, Discord, workers, and more, not raw SQL.

Get started

Up and running in minutes.

Clone the repo and run setup to install dependencies and open the desktop Host. Provider keys are optional, and your data lives in your per-user Cuttle home.

  • Windows 10/11 or Ubuntu 24.04+ (22.04 with a Python 3.11 install)
  • Python 3.11+ with python3-venv on Linux; Node.js 22.12+ for desktop
  • Install and authenticate your chosen agent CLIs separately: Cursor, Codex, Claude, Muse, OpenCode, …

For browser-only setup, add --browser on Linux or -Browser on Windows, then open https://127.0.0.1:8080.

Read the quick start Docs 
# clone and install
git clone https://github.com/kidchemical/Cuttle.git cuttle && cd cuttle
./setup_linux.sh

Your agents. Your machines. Your rules.

Get started Star on GitHub