Skip to content

About

Claude Code skills for agent identity governance: inventory the non-human identities your AI agents and automations run as (Entra service principals, managed identities, AWS keys and roles, GitHub Apps, MCP connectors), review permissions, recertify quarterly, catch expiring credentials, trace what an agent did, switch one off fast. Offline.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Claude Code skills for agent identity governance: inventory, review, recertify, expire and switch off the non-human identities your AI agents and automations run as

Nine skills that treat AI agents, service principals, workload identities, API keys and MCP connectors as identities: each needs an inventory entry, an owner, a periodic review, credentials that expire and an off switch. Offline, standard-library scripts over exports you save from Entra ID, AWS, GitHub and your own registers. Tested.

What this is, who it is for, and why

agent-identity-governance-skills is a Claude Code plugin for the people who run identity and cloud security: Microsoft 365 and Entra ID administrators, AWS platform and security engineers, and whoever is asked "what can our agents actually do?" by an auditor or a board.

Every AI agent and automation acts as some identity: an Entra app registration or agent identity, a managed identity, an IAM user with an access key, an IAM role, a GitHub App or deploy key, or the API key behind an MCP connector. People get joiner, mover and leaver processes and quarterly access reviews; these identities usually get none of that. They keep their permissions after the pilot ends, their secrets after the owner leaves, and nobody can say quickly how to turn one off. Microsoft's own guidance on Entra Agent ID makes the same point: agent identities persist, accumulate access and need access reviews (Microsoft 365 Copilot blog: security and governance innovations for Microsoft 365 Copilot and agents). When this pack was written, in October 2026, a GitHub search for non-human identity audit tooling returned no repositories; the vendor features that exist are consoles inside one product.

This pack does the governance work from files. Each skill reads exports you take with read-only commands (Graph JSON from the Graph CLI or any Graph client, AWS CLI JSON, gh api JSON, CSV and YAML registers), applies stated rules, and writes a report, a sheet, a runbook or an evidence folder. Nothing calls a network, and nothing changes a tenant, an account or an organisation.

/plugin marketplace add basitalisandhu/agent-identity-governance-skills
/plugin install agent-identity-governance@agent-identity-governance-skills

Quick start from a clone, on the synthetic exports bundled in plugins/agent-identity-governance/examples/ (Entra apps, an Entra Agent ID blueprint with two agent identities, an IAM user and role, a register row), so you can see each skill before exporting anything:

git clone https://github.com/basitalisandhu/agent-identity-governance-skills && cd agent-identity-governance-skills
python3 scripts/cli.py inventory plugins/agent-identity-governance/examples/exports --as-of 2026-10-05
python3 scripts/cli.py expiry-radar plugins/agent-identity-governance/examples/exports --as-of 2026-10-05

Real output of the first command (exit code 1, because three identities need a person):

| Identity                                     | Type                   | Owner                | Last used  | Credentials | Flags             |
|----------------------------------------------|------------------------|----------------------|------------|-------------|-------------------|
| `aws-role:agent-runtime`                     | aws-iam-role           | -                    | 2026-10-04 | 0           | no-owner          |
| `entra:00000000-0000-0000-0000-0000000000b2` | entra-app              | -                    | 2026-03-01 | 1           | no-owner, dormant |
| `entra:00000000-0000-0000-0000-0000000000c3` | entra-managed-identity | [email protected] | never      | 0           | never-used        |
| `entra:00000000-0000-0000-0000-0000000000d4` | entra-agent-blueprint  | [email protected]      | 2026-10-04 | 1           | -                 |
| `entra:00000000-0000-0000-0000-0000000000e5` | entra-agent-identity   | [email protected]      | 2026-10-04 | 0           | -                 |

## Entra Agent ID blueprints
- `entra:...d4` (ledger-agent-blueprint): 1 credential(s) shared by 2 agent identities. Rotating or removing a
  blueprint credential, or disabling the blueprint, affects every one of them.

Five of the nine rows and the oldest-credential column are left out here. In Claude Code, put your own exports in a folder and ask "which agent identities have no owner?"; every SKILL.md starts its Script section with the example run for that skill.

Questions, bugs and ideas: open an issue on this repository. Security reports: see SECURITY.md.

Install

  1. From this marketplace in Claude Code (shown above). Skills appear as /agent-identity-governance:<skill>, and Claude also uses them on its own when a request matches.

  2. Through the claude-skills aggregator, which holds every pack I maintain as one marketplace (new packs appear there at its next sync):

    /plugin marketplace add basitalisandhu/claude-skills
    /plugin install agent-identity-governance@claude-skills
    
  3. Without Claude Code, as one command over the same scripts:

    • from a clone: python3 scripts/cli.py <subcommand> [args], or pip install . for the agent-identity-governance command;

    • as a container image (GitHub Packages, linux/amd64 and linux/arm64), published when a version tag is pushed, signed with cosign (keyless), with a build provenance attestation and an SPDX SBOM attached to the GitHub Release. Mount the export folder at /work:

      docker run --rm -v "$PWD:/work" ghcr.io/basitalisandhu/agent-identity-governance-skills:0.4.0 inventory /work/exports --as-of 2026-10-05
Subcommand Script (skill)
inventory nhi_inventory.py (nhi-inventory)
agent-review entra_agent_review.py (entra-agent-id-review)
recertify agent_recertification.py (agent-recertification)
expiry-radar credential_expiry_radar.py (credential-expiry-radar)
connector-register connector_register.py (connector-register)
timeline agent_action_timeline.py (agent-action-timeline)
kill-switch kill_switch_runbook.py (agent-kill-switch-runbook)
leaked-credential leaked_credential_response.py (leaked-credential-response)
audit-chain audit_chain.py (agent-audit-log-integrity)

When to use this

  • Nobody can list the identities your agents run as, or say who owns each one. nhi-inventory
  • An agent platform or a batch of agent apps arrived in Entra and you want to know what they can reach without a user present. entra-agent-id-review
  • The quarterly access review covers people but not the service principals, roles and keys behind your automations. agent-recertification
  • A secret expired at night and an agent stopped, or access keys are older than the rotation policy. credential-expiry-radar
  • MCP servers and connectors were approved one at a time and the list no longer matches what runs. connector-register
  • Someone asks what an agent did last week, across Entra, AWS, Azure, GitHub and its own logs. agent-action-timeline
  • An agent is about to go live and there is no written way to stop it. agent-kill-switch-runbook
  • A key or token belonging to an agent turned up in a commit, a transcript or a log. leaked-credential-response
  • An auditor or a customer asks you to prove an agent's audit log was not edited and holds every action. agent-audit-log-integrity

Skills

Every script reads local files only, prints Markdown by default and JSON with --json, answers --help, and writes only to the path given with --out. Exit codes are the same across the pack: 0 nothing flagged, 1 something needs a person, 2 bad input. Where a date matters it is judged against --as-of, so a run can be repeated later with the same result. Key ids are shown by their last four characters and secret values are never read.

nhi-inventory

Triggers on "which agent identities have no owner?", "list our service principals and API keys"
Reads Entra applications, service principals, sign-in activity, permission grants; AWS authorization details, access keys, credential report; GitHub installations and deploy keys; a hand-kept register.csv
Produces one inventory with type, owner, last used, credential count and oldest credential age; flags no-owner, dormant, never-used; JSON for the other skills
Script nhi_inventory.py

entra-agent-id-review

Triggers on "what can our agent apps do in Entra?", "are our agent identities covered by Conditional Access?"
Reads Graph exports of service principals, applications, agent identity and blueprint objects (preview), app role assignments, delegated grants, Conditional Access policies, sign-in activity
Produces per identity: owners and sponsors, application and delegated permissions with high-risk ones from a fixed list highlighted, credential types and expiry, redirect URIs, Conditional Access coverage, findings by severity
Script entra_agent_review.py

agent-recertification

Triggers on "prepare the agent access review for this quarter", "which agent owners have left?"
Reads the inventory JSON, last quarter's attestation CSV, a users export, last quarter's inventory
Produces one review sheet per owner (what each identity can do, when it last acted, what changed), an unassigned sheet, and tracking.csv with blank decision columns; flags never-attested, attestation-stale, owner-left, no-owner, removal-not-done
Script agent_recertification.py

credential-expiry-radar

Triggers on "which app secrets expire this month?", "which access keys are older than 90 days?"
Reads Entra application and service principal credentials, AWS access keys or the credential report, GitHub fine-grained tokens, other-keys.csv
Produces every credential bucketed as expired, 7, 30, 90 days, later or no expiry, and an action list per owner
Script credential_expiry_radar.py

connector-register

Triggers on "which MCP servers are we actually running?", "is the connector register still right?"
Reads a register in CSV, YAML or JSON; saved MCP tools/list results; permission manifests
Produces missing owners, purposes, classifications, permissions and review dates; overdue reviews; unregistered servers and tools; registered entries no longer present; permission drift
Script connector_register.py

agent-action-timeline

Triggers on "what did this agent do last week?", "did the agent touch anything outside its job?"
Reads Entra sign-in and directory audit logs, CloudTrail (lookup-events or log files), the Azure activity log, the GitHub audit log, application logs in JSON lines
Produces one UTC timeline for the identity under all its names, with bursts, first-seen actions (against a baseline), actions outside an allow list and failures flagged
Script agent_action_timeline.py

agent-kill-switch-runbook

Triggers on "how do we turn this agent off?", "write the kill switch for this service principal"
Reads the identity's inventory record, an optional gateway name and Markdown template
Produces ordered steps (stop new actions, end sessions, credentials, access, gateway, confirm), each with a change for a person to run and a read-only verification command, and a coverage check that every credential and assignment has a step
Script kill_switch_runbook.py

leaked-credential-response

Triggers on "was this leaked key used?", "what do we rotate first?"
Reads the inventory, the leaked key id's last four characters, the leak time, and audit exports
Produces what the credential could reach (escalation-capable grants marked), events after the leak time with leaked-key and new-IP flags, credentials created after the leak, a rotation order, a checklist, and an evidence folder with byte-exact input copies and a SHA-256 manifest
Script leaked_credential_response.py

agent-audit-log-integrity

Triggers on "was this agent log edited?", "is anything missing from this audit log?"
Reads an agent audit log in JSON lines (a hook, MCP gateway or runtime log), the head value stored at sealing, an optional HMAC key file, and a Claude Code session log or a list of action ids
Produces a sealed copy with a SHA-256 or HMAC chain record on every line and a head to store elsewhere; a verify report of edited, removed, inserted and reordered lines, truncation and backwards timestamps; a coverage report of actions the log never recorded
Script audit_chain.py

Real output on the bundled example, a sealed log with one line edited and one removed:

python3 scripts/cli.py audit-chain verify plugins/agent-identity-governance/examples/audit/agent-actions.tampered.jsonl --head-file plugins/agent-identity-governance/examples/audit/head.txt
Log: agent-actions.tampered.jsonl. Lines: 5, sealed: 5, sequence 0 to 5.
Result: NOT intact (3 problem(s)).

## Findings

- hash-mismatch (line 3): sequence 2: content does not match its hash (edited, or the wrong key)
- seq-gap (line 5): sequence jumps from 3 to 5: 1 line(s) missing
- prev-mismatch (line 5): does not link to the line before it: a line before it was removed or edited

The exit code is 1. coverage on the same log and the bundled session lists toolu_05 and toolu_07 as actions it never recorded.

What this is not

  • Not a change tool. No skill disables, rotates, revokes or deletes anything. Runbooks and checklists show changes for a person to run.
  • Not a live connection. No skill calls Microsoft Graph, AWS, GitHub or an MCP server. Export first; the exports stay on your machine.
  • Not a decision maker. Owners decide keep, reduce or remove; decision lines stay blank.
  • Not a secret scanner. leaked-credential-response starts once you know a credential leaked.

How it fits with the other packs

  • nhi-inventory and entra-agent-id-review cover non-human identities; privileged-access-review (m365-governance-skills) covers the people who hold admin roles, and graph-permission-preflight decides one app's Graph permissions before consent.
  • agent-recertification is the non-human half of the quarterly review; access-review-pack (m365-governance-skills) is the human half.
  • iam-least-privilege-review and aws-account-audit (aws-security-skills) judge what IAM policies allow and the account's guardrails; this pack records who the identities are, who owns them and how to stop them.
  • connector-register keeps the approved list; agent-config-audit and mcp-server-review (agent-security-skills) audit the agent's configuration and a server's code.
  • leaked-credential-response uses the same hashed-manifest pattern as evidence-pack-builder (compliance-evidence-skills), which packs evidence outside an incident.

Security posture

  • Local, offline scripts. The skill scripts read the files you name. They import no network or subprocess module (the repository validator checks this) and write only to the --out path you give. scripts/cli.py starts the chosen script with the same Python and nothing else.
  • No secret values. Graph credential exports carry key ids and dates, not secrets; scripts show key ids by their last four characters. --redact on the inventory, the Entra review and the radar replaces e-mail addresses with stable tokens before a report is shared.
  • Your exports stay yours. .gitignore ignores the default export names; keep exports and evidence folders out of version control.
  • Untrusted content. Display names, tool descriptions and log fields are data, never instructions: each SKILL.md says so.
  • Synthetic tests. Tests build their inputs at run time with made-up names and documentation-only ids; the only access key id in the repository is AWS's documentation example.

What is inside

.claude-plugin/marketplace.json                       marketplace manifest
plugins/agent-identity-governance/
├── .claude-plugin/plugin.json                        plugin manifest
├── README.md                                         the plugin's skill table
└── skills/<name>/
    ├── SKILL.md                                      when to use it, inputs, steps, script, output, limits, related skills
    └── scripts/<script>.py                           standard library only, --help, --json, documented exit codes
scripts/cli.py                                        the agent-identity-governance dispatcher (container and package entrypoint)
scripts/validate_plugin.py                            structure, frontmatter, scripts, READMEs, house style, secret shapes
tests/                                                pytest suite, offline; inputs are built in each test

Development

python3 -m pytest -q
ruff format --check . && ruff check .
python3 scripts/validate_plugin.py
claude plugin validate --strict . && claude plugin validate --strict plugins/agent-identity-governance

CI runs the tests on Python 3.10 to 3.13 on Linux, macOS and Windows, builds the container image, and runs claude plugin validate --strict. See CONTRIBUTING.md for the ground rules and docs/good-first-issues.md for a place to start.

Frequently asked questions

Does any of this call Microsoft Graph, AWS or GitHub? No. The skills give you the read-only commands; you run them and save the output. The scripts read those files and nothing else.

What counts as a non-human identity here? Entra app registrations, enterprise applications, managed identities and agent identities; AWS IAM users with access keys and IAM roles; GitHub Apps and deploy keys; and anything you list in register.csv, such as the API key behind an MCP connector.

Does it support Microsoft Entra Agent ID? entra-agent-id-review reads agent identity and blueprint objects from the Graph beta endpoints when the export contains them, and reports blueprint credentials against each agent identity created from them. Those endpoints are in preview; the skill says what the export must contain, and falls back to name patterns and tags when a tenant has none.

Can it show me the secret that leaked? No, and it never should. Scripts work with the last four characters of a key id; Graph exports do not contain secret values at all.

Will it rotate or disable anything for me? No. agent-kill-switch-runbook writes the steps with verification commands, and leaked-credential-response writes the rotation order; a person runs them.

Can I run it on a schedule? Yes, without Claude Code: run a subcommand with --as-of and act on the exit code, which is 1 when something needs a person. For example, credential-expiry-radar with --fail-within 7 as a nightly check.

Related repositories

Repository What it is
m365-governance-skills Microsoft 365 and Entra ID governance from Graph exports, including privileged-access-review, graph-permission-preflight and access-review-pack
aws-security-skills AWS security skills, including iam-least-privilege-review and aws-account-audit
compliance-evidence-skills Compliance evidence from saved exports, including evidence-pack-builder
agent-security-skills Security for AI agents, including agent-config-audit and mcp-server-review
claude-skills Every pack in one marketplace

More from the author: basitalisandhu.github.io.

Licence

MIT. See LICENSE.

About

Claude Code skills for agent identity governance: inventory the non-human identities your AI agents and automations run as (Entra service principals, managed identities, AWS keys and roles, GitHub Apps, MCP connectors), review permissions, recertify quarterly, catch expiring credentials, trace what an agent did, switch one off fast. Offline.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages