Early access · An orchestrator for software projects
Discuss the vision. Lea runs the work.
Minga gives each project a lead agent. Talk through what you want with Lea. She turns it into milestones and items, works out what depends on what, hands the work to Claude Code and Codex on your machines, and follows through, as far on her own as you choose.
The lead
Goals and ideas, not agents and threads.
Most tools give you an agent and a thread to watch. Minga gives you a lead. Lea owns the plan: she orders the work by what waits on what, starts each item when its turn comes, reviews the result against the item's checks, and decides the next step. You decide how far she goes without you.
Today
- Goals with KPIs, milestones, items with checks for done, and dependencies between them.
- Ready items start on their own once their milestone is in Now; Lea reviews what comes back and sends it back when it misses the checks.
- Three levels: Suggest in chat, Ask first, Act, then tell me (the default). Switch any time.
- At every level, anything outward, a colleague's items and the brief wait for you. Pause stops everything.
- Agents can ask each other, on the record.
Coming
- A daily check-in where Lea reviews each goal against its KPI.
- Agent pages that show what each agent knows and is doing.
Where work runs
Run it anywhere. Move it in between.
Pair the machines you already have: your laptop, the Mac mini under the desk, a VPS. Each one reports what it can do, from the CLIs and logins on it to Xcode or Docker, and each item goes to a machine that has what it needs. When one goes to sleep, the next run goes to another that qualifies.
Use what you already pay for. Codex runs on your ChatGPT plan through the Codex CLI. Claude runs through your own Claude Code setup on your machine. Minga never holds your login.
Today
- The
mingarunner on macOS and Linux, laptop or VPS. It dials out; the cloud never connects into your machine. - Stock Claude Code and Codex, with your own settings, permission modes and MCP servers; one profile per CLI setup.
- Capabilities: runners report tools such as
xcode,simulator,dockerandgh; an Area can require them or pin a machine. - A run goes to any online machine that qualifies; if none is online, the item says it is waiting for a runner.
- Each item works in its own git worktree of your existing clone.
Coming
- Minga sandboxes: runners in Minga's cloud for work that needs nothing of yours but the repo, many at once, with API keys.
- Cloud runners: hand items to hosted agents: Claude Managed Agents with an Anthropic API key, Codex cloud on your ChatGPT plan through your runner.
- Windows runners.
Teams
One lead, one plan, the whole team.
Invite your colleagues to the workspace. Everyone works with the same Lea, the same plan and the same record of decisions, so the team shares one understanding of where things stand. Every message and change says who made it, and the machines Lea works on serve the whole team.
Today
- Invite with a link; roles Member, Admin, Owner.
- One Lea and one plan per project, shared by everyone; one chat with Lea, with a name on every message, and “via Claude Code” or “via Codex” when it came over MCP.
- Cards go to one named person; owners and admins can decide them too.
- Shared runners: Lea's work, whoever asked for it, runs on the machines of the person who looks after her.
- The Team tab shows every agent and worker, what it is on and on which machine; the plan shows each item's owner.
Coming
- Your own chats with Lea, readable by the team, and Lea's own log of her follow-ups.
- Presence, and For you / Team in Needs you.
- Hands follow the asker: your request runs on your machine or a team machine; each machine is Just me or The team.
- Anyone may decide any card, labelled “for Becky”.
Visibility
See the plan, not just the chat.
A chat thread tells you what was said. Minga shows you where the project stands: the plan, what waits on what, what is running and where, what came out, and what is waiting on you.
Plan
List by milestone, a Canvas of dependencies, and Outputs. Progress counts finished items only.
Team
Every agent and worker, what it is on, and on which machine.
Needs you
Approvals, answers and reviews in one place; pick between outputs side by side.
Pull requests New
Runners watch each item's pull request with your own
gh; Merge sits on the review card.
How it works
Step 1: Pair a machine
Install the
mingarunner and pair it. It finds your Claude Code and Codex setups and your repos.Step 2: Talk it through
Tell Lea what you want. She asks what she needs and writes the brief.
Step 3: Agree the plan
Milestones and items with checks for done. Moving a milestone to Now is the go-ahead.
Step 4: Decide what needs you
Approvals, answers and reviews collect in Needs you. Lea carries on from your decision.
curl -fsSL https://app.minga.run/install.sh | shInstall the runner (macOS or Linux, no sudo)minga runner loginPair it with your workspace
From your terminal
Talk to Minga from Claude Code and Codex.
Connect Minga once as an MCP server, and your own Claude Code or Codex session works with your projects as you. Starting from an existing repo? The import prompt drafts the brief, milestones and open questions from it, and writes nothing until you approve.
claude mcp add --transport http --scope user minga https://app.minga.run/mcpClaude Codecodex mcp add minga --url https://app.minga.run/mcpCodex/minga:import_repo_state dictlyThen, in Claude Code: draft the project from an existing repo
Questions
Can I use Minga today?
Minga is in early access, and things change weekly. Create an account in the app and pair a machine to start; if you already have an account, open the app.
How is this different from running Claude Code or Codex myself?
You still run them; Minga runs them for you. It adds the level above the session: a plan that outlives any thread, a lead who orders the work by its dependencies, reviews it against each item's checks and follows through at the autonomy you set, one place for what needs you, and work spread across your machines and your team. Lea and workers are instructed to open draft pull requests and leave merging to you; Merge is your button on the review card.
Which subscriptions and models does it use?
Your own. Codex runs on your ChatGPT plan through the Codex CLI; Claude runs through your own Claude Code setup on your machine. Minga does not sell or broker model access and never holds your login. Lea can think on one of your runners the same way, or in Minga's cloud with an Anthropic or OpenAI API key you add. When they come, Minga sandboxes will use API keys; cloud runners follow each vendor's terms: Claude Managed Agents an Anthropic API key, Codex cloud your ChatGPT plan through your runner.
Does my code leave my machine?
Your repo stays on your machine and work happens in your clone. What a run reads and does (its transcript: commands, output, file excerpts and diffs) is sent to Minga so you and your workspace can follow it, and the CLI sends your code to its model provider as it always does. Minga also stores the plan, brief, chats, history and attached outputs. Details: Concepts › Data and security.
Does my laptop have to stay on?
Only while work runs on it. Pair a VPS to keep things moving overnight; the next run goes to any online machine that qualifies. Work that only your laptop can do (an Xcode build, say) waits for it and says so.
Can my team use it?
Yes: invite people with a link, share one Lea and one plan per project, and Lea's work runs on the machines of whoever looks after her. Own chats per person, presence and runs that follow the asker are coming. See Invite your team.
What does it cost?
There is no pricing yet. During early access, model usage runs on your own plan or API key and is billed to you by that provider.
What isn't it for yet?
Windows runners, single sign-on, a self-hosted control plane, or hosted machines with nothing of your own to pair (Minga sandboxes are coming). Minga is not open source today.