New: LLM Rankings are live!

See how the top LLMs rank on real infrastructure tasks

Your AI agent does not need cloud credentials.

Coding agents like Claude Code can now plan and change infrastructure. The usual shortcut is an AWS key or a broad IAM role, which gives the agent the reach of whoever set it up. Massdriver gives the agent a scoped token and keeps the cloud credentials on its own side of the boundary.

What a cloud key in an agent actually grants

A cloud credential does not know who is using it. An agent holding an access key or an assumed role can create, change, or delete anything that role allows, in every account and environment the role reaches. When something goes wrong, the cloud audit trail shows the role. It does not show which agent, which session, or which prompt.

The key also travels. It sits in the agent’s environment, and from there it can land in a context window, a transcript, a log line, or the payload of a prompt injection. Every place it lands is a place you have to rotate it from.

Narrowing IAM for each agent and each environment is possible. It turns into a second permission system that the platform team maintains by hand, in a policy language that has no concept of a production database owned by the payments team.

Common ways to scope an agent’s cloud access

Every option except the last still hands the agent a cloud credential. They differ in how much that credential can reach.

The developer’s own profile

What it gives you

Zero setup. The agent works immediately.

What it leaves to you

The agent has everything the developer has, and its actions look like the developer’s in the cloud logs.

A dedicated IAM role per agent

What it gives you

A separate identity in the cloud logs and a policy you can narrow.

What it leaves to you

IAM policies per agent, per account, and per environment, kept in sync by hand as resources come and go.

Short-lived credentials through OIDC or STS

What it gives you

Credentials that expire on their own, which limits a leak.

What it leaves to you

While it is valid, the credential carries full cloud-level permissions. It cannot express “propose in production but do not deploy.”

A separate cloud account per agent

What it gives you

Hard isolation. Nothing outside the account is reachable.

What it leaves to you

No path for the agent’s work to reach production, and more accounts to govern.

A scoped Massdriver token

What it gives you

The agent holds no cloud credential. Policy on environment, team, and blast radius decides what the token may do, and every action is recorded with the agent as the actor.

What it leaves to you

You connect your cloud to Massdriver once and write the group policies for your agents.

How the access is split

01

The agent holds a Massdriver token

The token belongs to a service account, with an expiry you set. The agent receives it through the Massdriver MCP server’s environment and never sees a cloud key.

02

Policy decides what the token may do

The service account belongs to a group, and the group’s attribute-based policies say which actions it may take on which environments, projects, or teams.

03

Massdriver holds the cloud credentials

AWS, Azure, and GCP credentials stay in Massdriver, which uses them only for deploys the policy allows.

04

Production needs a person

Proposing a change and deploying it are separate permissions, so an agent can propose a production change while a person with deploy permission decides.

One agent, from setup to revocation

A platform engineer sets up access for a team’s coding agent. A week later the agent needs a change in production.

  1. 01
    What happens

    The platform engineer creates a service account for the agent, with a token that expires in 30 days, and adds it to the coding-agents group.

    What the record shows

    The service account is its own actor type, so its actions are never mixed in with a person’s.

  2. 02
    What happens

    The group’s policy lets it deploy in dev and staging, and only propose in production.

    What the record shows

    Massdriver shows each policy in plain language, so a reviewer can read what the group may do.

  3. 03
    What happens

    The agent changes a staging instance. The policy allows it, and the deploy runs with Massdriver’s credentials.

    What the record shows

    The deploy is recorded with the agent’s service account as the actor.

  4. 04
    What happens

    The agent needs the same change in production. A deploy there is not allowed, so it submits a proposal.

    What the record shows

    The proposal waits, with its plan diff and policy results.

  5. 05
    What happens

    An approver with deploy permission in production reviews the proposal and approves it.

    What the record shows

    The approver and the proposer are both on the record.

  6. 06
    What happens

    The team stops using the agent, and the engineer revokes the token.

    What the record shows

    Access ends at once. No IAM role, key, or cloud policy has to change.

Write permissions for kinds of resources

Agents create, replace, and destroy resources far more often than people do. A permission tied to a specific resource ID breaks every time that resource is recreated. Massdriver policies are written against attributes such as environment, team, and blast radius, so a rule like “this group may manage non-production databases” keeps holding while the resources underneath it change.

The rules are hierarchical, so they cascade down through projects and environments, and they sit above every cloud you run. Cory O’Daniel explains why the unit of permission has to change in Your Golden Paths Weren’t Built for Agents.

Start with agents that propose

The simplest first policy lets agents deploy in dev and staging and only propose in production. Agents get a fast loop where mistakes are cheap, and every production change has a person behind it.

From there you widen access by changing grants and blast-radius attributes, environment by environment. An agent that has proposed a hundred clean changes to low-risk services can earn deploy permission for those services, while databases and networks stay behind an approver.

Questions security and platform teams ask

Are short-lived cloud credentials enough?

They limit how long a leaked credential is useful. While it is valid, it carries full cloud permissions, and it cannot separate proposing a change from deploying it. A Massdriver token keeps the agent out of the cloud entirely.

Can the agent read production?

If you grant it. Read access to the catalog and the graph is a separate permission from changing anything, and you set it by environment.

What do our cloud audit logs show?

The cloud sees Massdriver’s credentials. Massdriver’s own audit log names the agent’s service account on every proposal, approval, and deploy, and that record can be exported.

How do we cut off an agent?

Revoke its token or remove its service account from the group. Its access ends with no change to cloud IAM.

Which agents does this work with?

Claude Code today, through the MCP server and the Architect plugin. Any agent that can use the MCP server gets the same scoped access.

When a per-agent IAM role is enough

If one agent works in one sandbox account, with no path to production and no shared data, a narrow IAM role with short-lived credentials covers the risk. Scoped tokens matter when agents work in the same accounts as your services, and their changes have to reach production.