A deploy nobody approved
A pipeline holds production credentials. A change merged late on a Friday ships, because the credential worked. Nobody decided that it should.
Start here
Why AtlaSent matters, how it helps, and how to start.
A short guide for anyone new to AtlaSent: the problem it solves, what changes once it is in place, the three ways to start, and what you get on each.
Why it matters
Being able to act is not the same as being allowed to.
Your systems check who someone is and what they can reach. They rarely check whether your organization agreed to this exact action, right now. That gap is where costly mistakes happen.
An agent with a key
An AI agent given an API key to help with support issues a refund or deletes records, because the key allows it. The agent was authenticated. It was never authorized to do that.
An approval that drifted
A reviewer approves one commit. The pipeline later deploys a different one under the same approval. The ticket says approved, but not for this change.
Illustrative situations, not customer incidents. In each one, every identity and access check passed.
How AtlaSent helps
One check, at the moment the action runs.
- Ask first. Before the deploy, the refund or the tool call runs, your system asks AtlaSent.
- Check your rules. Who may do this, who must approve, which duties stay separate, and under what conditions.
- Allow, hold or refuse. If a person must approve, the action waits. If it is not allowed, it does not run. If AtlaSent cannot decide, it does not run either.
- Only that action. An allowed action gets a single-use permit for exactly that request. Change the commit, the target or the amount and it needs a new decision.
- Keep the proof. Every decision, including every refusal, leaves a signed record of what was asked, by whom, and why it was decided.
The three situations, with AtlaSent
The Friday deploy is held until the person your rules name approves it. Without that approval the pipeline step fails and nothing ships.
The agent’s refund is checked like any other action. Over the limit your rules set, it waits for a person.
The swapped commit does not match the permit, so it is refused and needs its own approval.
How to start
Three ways in. Pick the one that fits.
1. Try it free, on your machine
For a developer who wants to see it work today, with no account.
- In Claude Code, add the agent guard:
/plugin marketplace add AtlaSent/atlasent-mcp-server, then/plugin install atlasent-guard@atlasent. - Ask your agent to run something destructive. It waits for you instead.
- Prefer other MCP hosts? MCP Gate puts the same fail-closed check in front of their tools.
2. Gate one deploy
For a team that wants its own production deploy to ask first.
- Create an account and an API key. The free key covers a starter allowance.
- Save it as the
ATLASENT_API_KEYsecret in your repository. - Add one AtlaSent step before your deploy step, and run the deploy only when it returns
verified. - Want risky deploys to wait for a teammate? That is Team Cloud.
3. Protect one action across the organization
For a security, platform or risk owner who needs it to hold up in front of an auditor.
- Pick the one consequential action you most need to control.
- With us, map the route that performs it today, including any way around.
- Set your requirements, then prove a refusal and a valid run on that route.
- Switch production protection on when you authorize it. This is Authority Launch.
What you get
What each path gives you.
Free
- Local blocking on your machine: the Claude Code agent guard and MCP Gate. Local blocking never requires payment.
- A real API key with hosted decisions and signed receipts, within a monthly free allowance.
- The TypeScript and Python SDKs, the REST API, the GitHub Action and the MCP server.
Team Cloud
- Shared policies, in one place for the whole team.
- Hosted approvals: risky actions wait for the right teammate, from any pipeline or agent.
- Members, roles, API keys and environments managed in the console.
- Hosted decision history and signed receipts for the team.
Enterprise
- Protection across teams and systems, starting with one action proven in production.
- Every known route to a protected action shown as Complete, Partial, Unknown or Not configured. Unknown is never shown as protected.
- Distinct identities for people, pipelines and AI agents, and separation of duties your administrators cannot quietly override.
- Evidence you can hand a reviewer, and implementation and support written into the agreement.
What AtlaSent does not do
Honest limits.
- It does not replace your identity, pipeline or approval tools. It uses them, and adds the decision between approval and execution.
- It only decides actions that pass through it. A route that goes around the check is not protected, and AtlaSent shows it as a gap rather than hiding it.
- It does not judge content. It decides whether your organization authorized the action, not whether a prompt or a message is safe.
AtlaSent Enterprise
Ready to protect your first action?
Start from the one consequential action you most need to control. We look at it and the route that performs it today.