grantry 0.3.0

Local credential broker for humans and AI agents, AWS-first

grantry

A local credential broker for humans and AI agents. AWS first.

grantry logs you into AWS IAM Identity Center once, then hands out short lived credentials on demand. You get a clean command line. Your AI coding agents get the same credentials over MCP, but only for the accounts and roles you allow, only for as long as you permit, and every request is written to an audit log.

Everything stays on your machine. grantry talks to AWS and nothing else. No account, no server, no telemetry. Tokens live in the OS keychain, never in a plain file, and no secret is ever written to a log.

---

Why grantry exists

Every other credential tool assumes a person is at the keyboard. But agents are now heavy users of cloud credentials, and they are bad at logging in. They cannot click a device flow link, and they stall when a session expires. The common workaround is pasting long lived keys into the agent's environment. Those keys never expire and are never audited. That is a real security problem.

grantry removes the workaround. The agent asks, grantry checks your rules, and hands over short lived credentials or a clear refusal. You keep one login, one policy, and one audit trail across every agent on your machine.

Install

uvx grantry --help        # run without installing
# or
pipx install grantry

Or from source:

git clone https://github.com/saimeda32/grantry
cd grantry
uv sync
uv run grantry --help

Works on macOS, Linux, and Windows. Python 3.10 or newer.

Quick start

# 1. Point grantry at your Identity Center, just once. It remembers.
grantry login --start-url https://your-org.awsapps.com/start --region us-east-1

# 2. See what you can reach.
grantry ls

# 3. Generate a starter policy from your real access, then edit it.
grantry init

# 4. Run any command as a role.
grantry run my-dev/AWSReadOnlyAccess -- aws s3 ls

After grantry login, the native aws CLI, boto3, and Terraform work too. Run grantry populate once to create the matching profiles in ~/.aws/config, then use aws --profile ... with no grantry in the loop.

Route native tools through grantry (audited)

If you want every native credential fetch to go through grantry, so it is checked against your policy and written to the audit log, add a credential_process profile instead of a plain SSO profile:

[profile prod-readonly]
credential_process = grantry credential-process --identity prod/AWSReadOnlyAccess
region = us-east-1

Now aws --profile prod-readonly ..., boto3, and Terraform all get their credentials from grantry. This is also how you make grantry a real boundary for a sandboxed agent: give the sandbox only a credential_process profile with --caller agent, and the agent cannot reach anything the policy denies.

Commands

Command	What it does
grantry login	Log in to Identity Center once, for all accounts and roles.
grantry ls	List the account and role identities you can use.
grantry run <id> -- <cmd>	Run a command as a chosen identity.
grantry switch [id]	Print shell exports to adopt an identity. Omit the id to pick interactively.
grantry console [id]	Open the AWS console in your browser as an identity. Omit the id to pick.
grantry credential-process --identity <id>	Emit credentials as JSON for an AWS config credential_process entry, so native aws/boto3/Terraform route through grantry.
grantry populate	Write ~/.aws/config profiles for your access. Adds, updates, and prunes only its own profiles.
grantry check	Diagnose your session and access, with clear exit codes.
grantry init	Generate a working policy from your real access.
grantry audit	Print the grant history, or write an HTML timeline with --visualize.
grantry graph	Write an HTML map of what your agents can reach under the policy.
grantry mcp	Run grantry as an MCP server for agents.
grantry install [client]	Add grantry to an AI client's MCP config. Auto detects all if none named.
grantry admin assignments --as <id>	Crawl who has what across the whole org. Admin only. Add --snapshot to save it, or --diff to see what changed since the last snapshot.
grantry logout	Clear the saved session for the current instance.
grantry instances / grantry use <name>	List remembered orgs, or switch between them.
grantry install / grantry uninstall	Add or remove grantry from an AI client's MCP config.
grantry version	Print the version.

Use it with your AI agents

One command wires grantry into your AI clients:

grantry install            # auto detect every client you have
grantry install cursor     # or a specific one
grantry install --dry-run  # preview without writing

Supported: claude-code, claude-desktop, cursor, windsurf, vscode. grantry is added without touching your other MCP servers, and each client gets its own audit label. Restart the client to load it.

The agent then has four tools: whoami, list_identities, get_credentials(identity, ttl), and check_access(identity). If no one is logged in, the agent can call request_login, which notifies you and waits for your approval, then resumes on its own.

Policy

Write ~/.grantry/policy.yaml, or let grantry init generate it from your real access. See examples/policy.yaml.

agents:
  allow:
    - identity: "*/AWSReadOnlyAccess"
    - identity: "dev-*/AWSPowerUserAccess"
  deny:
    - identity: "*prod*/*Admin*"
  max_ttl: 15m
humans:
  max_ttl: 12h

Three rules govern it:

1. A deny always beats an allow.
2. For agents, anything not allowed is refused. Safe by default.
3. For you, anything not mentioned is allowed.

An identity is account-name/role-name, and * is a wildcard, so dev-*/AWSReadOnlyAccess means read only in any account whose name starts with dev.

A note on TTL and AWS: grantry cannot shorten an SSO credential below the lifetime AWS issues it with, because the reserved SSO roles do not allow client side re assumption. grantry reports the real AWS expiration and adds an advisory when a credential outlives your policy cap. The real control for short sessions is the permission set session duration, set by an admin in IAM Identity Center.

Admin: see who has what across the org

grantry admin assignments --as your-mgmt/AWSAdministratorAccess --visualize

This crawls the whole organization and writes an interactive graph of principals, permission sets, and accounts, with the links between them. It is safe to offer because AWS is the gatekeeper: only an identity that can assume a management or delegated admin role gets any data. The crawl caches principal names and uses retry hardening, so it handles organizations with thousands of assignments.

How your data is stored

grantry uses no database. It is a single user local tool.

• Secrets (SSO tokens) live in the OS keychain, through keyring.
• State lives as plain files in ~/.grantry/: instance.json, policy.yaml, and an append only audit.jsonl (mode 0600).
• Interop: grantry login also writes the AWS CLI token cache in ~/.aws/sso/cache/, the same file aws sso login writes, so the native tools work.

Everything survives reboots. To remove grantry state: grantry logout and rm -rf ~/.grantry.

Security

• Secrets (SSO tokens, refresh tokens) live in the OS keychain. Logging redacts tokens in one place, including exception tracebacks.
• The MCP server is not a network service. It talks to the agent that started it over stdio.
• Every grant is recorded in ~/.grantry/audit.jsonl, and never includes the credentials themselves.

Read this honestly before relying on grantry as a control: the policy gate only covers the MCP door. If an agent also has a shell and you have run grantry populate, the agent can use aws --profile ... directly and bypass the gate. For the gate to be a real boundary, run the agent with no ambient AWS access. And get_credentials returns credentials as text into the agent's context. See SECURITY.md for the full picture and to report a vulnerability.

Roadmap

grantry v1 covers AWS Identity Center. The provider layer is written so Azure and GCP can be added without touching the engine, policy, audit, or MCP surface. Team mode (shared, signed policy) comes after that.

Contributing

See CONTRIBUTING.md. In short: fork, branch, write a test, keep ruff, mypy, and pytest green, open a pull request.

License

Apache 2.0.