A free, open-source meta-harness for your coding agents.
Free & open source · Mac & Windows.Start your first task
Models, subscriptions and local runtimes
OpenAI
Anthropic
Gemini
xAI
Meta
DeepSeek
Qwen
Mistral
Kimi
Z.ai / GLM
Thinking Machines
NVIDIA
Cohere
MiniMax
ByteDance Seed
Poolside
Liquid AI
ERNIE
Hunyuan
Xiaomi
HyperCLOVA
Sakana AI
OpenRouter
Claude subscription
Cursor
GitHub Copilot
Windsurf
AntigravityInteractive examples, not live runs or benchmarks.
Seats draft alone, then talk in turn until the chair writes up the answer.
Opuschairdrafting…GPT-5drafting…Geminidrafting…Kimidrafting…Drafting aloneStream or buffer the parser input?
Each seat answers alone, no tools, before anyone hears the others.
Every seat bills itself; 4 seats make 13 model calls. Only the answer lands in the thread.
Steps run in parallel or in sequence; each waits only on what it depends on.
1Read the project
Implement, test and document together; review, then return the result.
Full end-to-end context observability, including where your tokens are being used.
Summarize 48k of earlier messages and results into an 8k handoff.
Summaries lose detail. Saved notes remain separate.
Select a span to inspect it.
Implements while another subagent reviews.
Spans overlap; durations don’t add up. Long traces may be shortened.
Every write, command and click asks first. The default.
File writes and searches go through. Commands, the browser and the pointer still ask.
Everything runs on its own except ten prohibited actions, like deleting outside what you named or sending your data elsewhere. Anything that matches stops and asks you. The screening prompt and the model that runs it are both yours to change.
Every tool runs on its own except computer use, which always asks. Other apps still ask. During computer use, global Escape stops computer access while the agent continues.
Inspect the evidence, ask Shinbo for a scoped change, and compare it on a replay before deciding to keep it.
How learning worksRun Claude Code, Codex and other CLIs in one thread.
Delegate work; inspect progress and results.
Steps run in parallel or in sequence; each waits only on the steps it depends on.
APIs, subscriptions, Ollama, LM Studio and llama.cpp.
Control edits, commands and app access.
Threads, goals, terminal, browser and artifacts. Quick Ask and voice on Mac.
Keep Markdown notes in a folder or Obsidian vault.
Schedule agents, scripts and conditions.
Inspect failures, compare a repair, and decide what to keep.
Add MCP servers, skills, plugins and generated UI.
Select your operating system below.
Choose a downloadPublished installers are available for Apple silicon Macs and Windows x64.
Choose your system to see the right installer.
Direct downloads
Shinbo v0.9.2. Release notes & checksums
Setup requires an OpenRouter key (a free key works), even with a subscription. Sign into your agent CLIs; providers bill separately. Setup guide
No Intel Mac or Linux installer. Mobile unreleased. Mac-only features may differ on Windows.
No. Setup requires an OpenRouter key; a free key works. Paid models and optional subscriptions bill separately.
Yes. Install and sign into supported agent CLIs, then use them from Shinbo. Provider access and usage limits still apply.
Local models are supported, but setup, secondary models, web tools and integrations may still need the internet.
Threads, plans, traces and notes stay on your computer. Prompts, relevant history and tool results are sent to your selected model service; integrations use their configured services.
shinbo
Linear
Notion
Figma
Obsidian
Slack
Vercel
Stripe
Sentry
Supabase