SecurityTrust model
Your code.Your keys.Your call.
Mamba is built so the safe path is the default one: local-first, isolated by git, and unable to merge, push or deploy without you.
Principles
Safe by default, not by setting.
Runs on your machine
The runner lives on your laptop or server and works on your repositories where they are. Your code does not pass through a Mamba service.
One worktree per task
Each task gets its own git worktree on its own branch. Agents never edit your checkout, and two tasks never share files.
Permissions come to you
When an agent asks to run something your rules do not cover, it waits for you. Pushes, deploys and deletes outside the repo always ask.
No model billing
Agents run on your own subscriptions and keys. Mamba never sits between you and a model provider's bill.
An audit trail by design
Every command, prompt, answer and checkpoint is an event in the log. The audit page reads straight from it.
Scoped keys
Other AI agents get their own keys to drive Mamba over MCP or REST. Invited viewers can read, not act.
Approvals
Honest permissions.
A permission prompt shows exactly what the agent wants to run and where. You allow it once or deny it. Automation only answers what your own rules cover.
- Allow once is the answer on the board. Standing approvals are rules you write, not a button you hit by accident.
- Automatic approval is a runner setting you turn on, and the board says so.
- Outside the repo, the answer always comes from you.


What Mamba never does
Lines we do not cross.
Scrape a terminal
If an agent has a structured protocol, that is the only way Mamba talks to it.
Claim unreported work
Status comes from what the agent and git reported. Never from a guess.
Merge on its own
A task can be ready to merge. Only you merge it.
Hold your model keys
Agents sign in with their own accounts, on your machine.
Found a security issue? Please report it privately to the maintainers before opening a public issue.