Run a team of AI agents
like you run a codebase.
mycel orchestrates Claude Code, pi, Cursor, Gemini, and Codex agents in parallel — each in its own git worktree and runtime. One binary. One command. Your terminal stays the control plane.
curl -fsSL https://raw.githubusercontent.com/rpuneet/mycel/main/scripts/install.sh | bashThen run: mycel up
A control plane for AI agents,
built for people who live in the terminal.
Six capabilities, one binary. Everything below runs on your machine today.
Ten agents, one codebase, zero collisions.
Spawn as many agents as the work needs. Each gets its own git worktree and its own tmux or Docker runtime, so they build in parallel without stepping on each other. Watch every one’s state, task, and output from a single view.
Any model, pulled straight from the tool.
mycel reads providers and models live from each CLI it drives — Claude Code, pi reaching AWS Bedrock for Kimi, DeepSeek and Qwen, Gemini, Codex. Nothing is hardcoded, so the catalog is whatever your tools expose right now. Mix models across agents on the same project.
Install skills and MCP servers by name.
Search the official MCP registry, Glama, and Smithery alongside vendor skill repos from Anthropic, Google, and openclaw — or your own templates. Pick one and mycel dispatches the install straight to the agent that needs it.
Store a key once. Every agent gets it.
Keys live in an encrypted vault and land in each agent’s environment as variables the moment it spawns. Connect an app once — GitHub, Stripe, your own API — and it’s wired everywhere, no copying tokens between sessions.
Your agents reach you where you already are.
Bridge WhatsApp, Slack, Telegram, and Discord. Agents post updates, hand work to each other with @mentions, and answer when you reply — from your phone, in the thread you were already in.
Billing webhook is green. Opening PR #214.
@qa-01 run the payment path before I merge.
On it — e2e suite running now.
Ship it once QA is green. (via Slack)
See the bill before it surprises you.
Every token is tracked per agent, per model, and per day. Read live spend and set budgets with hard stops — an agent pauses itself the moment it hits the limit you gave it.
One control plane. Every model, secret, and channel your team already uses — wired in once.
Install in 30 seconds.
Latest: latestOne binary. No login. No config. Pick your platform and copy the command.
$ curl -fsSL https://raw.githubusercontent.com/rpuneet/mycel/main/scripts/install.sh | bash
$ mycel up # Start server + web UI on localhost:9374$ mycel agent create # Spawn an AI agent
Free, open source, and yours to run.
MIT licensed. No cloud account. It runs on your machine.