Know what your agents can do. Before they do it.

ControlLoop is a local-first scanner and deterministic CI gate for what your AI agents can actually do. It reads the files that define an agent, builds one graph of its capabilities, compares it to the state you approved, and fails the build when the agent's authority grows past the ceiling you declared. It never runs your code and never sends your source anywhere.

Watch: the gate in a minute · Read: the example walkthrough

ControlLoop CLI

An agent's real authority is spread across an agent card, an MCP config, an OpenAPI spec, an IAM policy, and, for most agents, the framework code itself: a CrewAI crew, a LangGraph graph, an OpenAI Agents SDK handoff chain. The same goes for the coding agents that work on your repository. Claude Code, Codex, Gemini CLI, Cursor, Kiro, Roo Code, and Cline are configured by committed files, and so are the workflows that run an agent in CI. No single file tells you that one pull request just gave an agent the ability to move money, or to run a script on every session start.

How it works. ControlLoop walks the repository and reads only committed files. Adapters turn them into an Agent Capability Bill of Materials (ACBOM): agents, tools, resources, identities, and permissions, with evidence for every node and edge. The graph is diffed against the baseline you approved, classified by a material capability taxonomy, judged by 24 policy rules and the capability ceiling you declared, and the verdict is written to an evidence artifact. No model decides anything; the same inputs always produce the same verdict. Secrets, prompts, environment values, and file contents never reach the output.

Development preview. init, bom, diff, scan, gate, rules, and example work and are covered by tests. Releases are published to PyPI. A container image on GHCR and a Homebrew formula are published by the release pipeline from the next tagged release; see Container image and Homebrew.

Features

What it reads

• 🗂️ A2A agent cards, MCP server declarations for eleven agent hosts, OpenAPI documents, and AWS IAM policies in JSON or Terraform. See the adapter table.
• 🧩 Framework code, parsed and never imported: CrewAI, LangGraph, and the OpenAI Agents SDK. A tool it cannot resolve is reported as the worst reachable capability, never dropped.
• 🤖 Coding agents' own configuration: Claude Code settings, hooks, skills, subagents, plugins and marketplaces; Codex; Gemini CLI; Cursor; Kiro; Roo Code and Cline through their MCP configuration; and the GitHub workflows that run an agent.

What it tells you

• 🎯 The blast radius, first. Every report opens with the single worst thing the agent can do right now: ! WORST REACHABLE financial-action: … can invoke Issue a refund. Material capability taxonomy.
• 🗣️ A plain-language authority statement for every change to an agent's configuration: This change gives Claude Code the power to run node .github/setup.js on every session start. Authority statement.
• 🔀 A capability diff, not a file diff: what became newly reachable against the baseline you approved, by class. Capability diff.
• 📏 A capability ceiling that blocks authority nobody anticipated, through routes no rule matches, and cannot be suppressed. Capability ceiling.
• 📚 24 policy rules with a plain message, a remediation, and a stable documentation link on every finding. Policy rules.
• 🧾 An evidence artifact on every gate run, including the failing ones, plus JSON and SARIF 2.1.0 output. Evidence artifact format.

How it behaves

• 🔒 Local and offline. Core scanning works with the network disabled. Nothing is executed: no MCP server is started, no endpoint is contacted. Privacy, security, and limitations.
• 🧮 Deterministic. Same inputs, same pinned versions, same digest, same verdict. Unknown is ranked worse than known, never safer.
• 🙈 Nothing protected leaves. Source, prompts, credentials, environment values, and customer data never appear in output, logs, errors, identity material, or artifacts.
• ⚙️ CI-native. A GitHub Action and a pre-commit hook run the same gate.

Note: the agent hosts whose project configuration is read are Claude Code, Codex, Gemini CLI, Cursor, and Kiro, each pinned to its documentation as read on a stated date (controlloop --version prints it). MCP declarations are also read from VS Code, Roo Code, Cline, Continue, Windsurf, and opencode.

Quick start

Installation

Install the CLI from PyPI with uv:

uv tool install controlloop
controlloop --version

pipx install controlloop works too. Without Python, the shell installer downloads and verifies the standalone binary for your machine:

curl -fsSL https://raw.githubusercontent.com/controlloop-security/controlloop/main/install.sh | bash

Every install path, with its verification step, is in Installation. To run from source instead:

git clone https://github.com/controlloop-security/controlloop
cd controlloop
uv sync --locked --all-groups
uv run controlloop --version

ControlLoop 0.1.1
Supported OASF versions: none yet
Agent host key sets (documentation read on): claude-code 2026-10-17, codex 2026-10-17, cursor 2026-10-17, gemini-cli 2026-10-17, kiro 2026-10-17, roo-code 2026-10-29
Rule pack: controlloop-community 0.9.0 (24 rules)

Requirements

• Python 3.12
• uv
• Git, for the automatic merge-base baseline that diff and gate use
• macOS or Linux. Windows is unsupported.

First run

Point it at a repository containing an agent:

uv run controlloop scan /path/to/your/agent-repo

Nothing to try it on? The public reference agent is a realistic, deliberately imperfect agent built for exactly this:

git clone https://github.com/controlloop-security/controlloop-reference-agent
uv run controlloop scan controlloop-reference-agent/scenarios/order-support-agent

! WORST REACHABLE  financial-action: widgetworks-order-support-agent can invoke
  Issue a refund

✓ PASSED

• material change: not evaluated

! MEDIUM   policy.undeclared-tools: Tool 'Look up an order' is reachable but not
  declared in controlloop.yaml.
    at agent-card.json

! MEDIUM   policy.undeclared-tools: Tool 'Reply to customer' is reachable but
  not declared in controlloop.yaml.
    at agent-card.json

! MEDIUM   policy.undeclared-tools: Tool 'Issue a refund' is reachable but not
  declared in controlloop.yaml.
    at agent-card.json

Suppressions: none declared

Exit code 0. New here? uv run controlloop example prints a six-step walkthrough that matches the example walkthrough guide.

The commands

scan is what you run at your desk. gate is what you run in CI. Every command and flag is in the CLI reference.

Screenshots

Every block in this README is real output, recorded against the public reference agent with the version above. Nothing is staged.

A pull request adds a tool the agent's ceiling prohibits. The blast radius got worse, the ceiling breach blocks, and the evidence artifact is written:

! WORST REACHABLE  financial-action: widgetworks-order-support-agent can invoke
  Issue a refund

✕ DEPLOYMENT BLOCKED

• GAINED  financial-action: widgetworks-order-support-agent invokes
  customer-payment-methods
• GAINED  sensitive-data-access: widgetworks-order-support-agent invokes
  customer-payment-methods
• GAINED  write-access: widgetworks-order-support-agent invokes
  customer-payment-methods

✕ CRITICAL policy.capability-ceiling: agent 'widgetworks-order-support-agent'
  reaches 'customer-payment-methods', which is a prohibited resource or external
  destination under its capability ceiling
    at agent-card.json

controlloop gate on scenarios/order-support-agent with a branch that adds customer-payment-methods. Exit code 1.

A pull request adds a Claude Code hook. The gate says what that lets the agent do before anything else:

✕ DEPLOYMENT BLOCKED

Authority:
  • This change gives Claude Code the power to run `node .github/setup.js` on
    every session start (`.claude/settings.json`, hook `SessionStart`). No
    approval is recorded.

• GAINED  code-execution: Claude Code invokes SessionStart command hook

✕ HIGH     policy.agent-hook-command-change: Added: Hook on event 'SessionStart'
  for host 'Claude Code' runs `node .github/setup.js`.
    at .claude/settings.json

controlloop gate on scenarios/storefront-agent-config, branch scenario/session-start-harvester-hook. Exit code 1. The GitHub Action posts the same statement as one pull-request comment.

Examples

The common uses, each with the real command and its real output.

Scan a repository

uv run controlloop scan .

The report above: blast radius first, then the verdict, the material change against the baseline, every finding with its rule and evidence path, and the suppressions in force. How to read each line is in Terminal report format.

See what a pull request changed

uv run controlloop diff . --baseline main

CONTROLLOOP / CAPABILITY DIFF

• BASELINE   explicit git revision: revision main (4c258dc876245856f97786c05f51157ee68e8e25)
• SCOPE      full -- nodes, edges, and effective authority

• ADDED    node        clp:tool:sha256:ccc68b72…  customer-payment-methods
• ADDED    edge        clp:edge:sha256:0234aa41…
• ADDED    capability  clp:agent:sha256:2321fe51…  invokes  clp:tool:sha256:ccc68b72…

3 changes: 3 added

Without --baseline, diff uses the committed .controlloop/baseline.json if there is one, else the merge-base with your default branch. diff <baseline.json> <current.json> compares two ACBOM files directly.

Gate in CI, with evidence

uv run controlloop gate . --default-branch main \
  --evidence-path evidence.json --sarif-path controlloop.sarif

gate returns one of four verdicts and always writes the evidence artifact, including on a failing run, which is exactly when you need it:

gate needs real git history for an automatic baseline, so check out with fetch-depth: 0. The full exit-code table is at the end of this file.

Catch a change to a coding agent's configuration

The second screenshot above. Any change to .claude/settings.json, .mcp.json, .codex/config.toml, .gemini/settings.json, .cursor/hooks.json, .kiro/, a plugin, a skill, a subagent, or a workflow that runs an agent produces an authority statement. Same name, different package? That is a changed grant too:

This change gives Claude Code the power to start MCP server `catalog-docs` by
running `npx -y @widgetworks-example/catalog-docs-mcp@1.5.0`, whose declaration
or tool list changed (`.mcp.json`, server `catalog-docs`). No approval is
recorded.

From scenario/mcp-package-drift, where catalog-docs moves from 1.4.2 to 1.5.0 under the same name. The limits of what the scanner can and cannot see are in Agent authority limits.

Tell ControlLoop what your tools do

This is the step people miss. Discovery reads your agent card and your OpenAPI spec, but nothing in a repository states that issuing a refund moves money. Without that, every capability is unclassified and every scan reports no material capability is reachable. Declare it in controlloop.yaml:

schema_version: "1.1.0"
owner: support-eng

tool_capabilities:
  - tool: Issue a refund        # the DISCOVERED tool name, not its id
    financial_action: true
    read_only: false
    sensitivity: confidential
  - tool: Purge ticket history
    destructive: true

Match the name discovery found, which for an A2A skill is its card name, not its id. Run controlloop bom and read the node names if you are unsure; a declaration naming a tool that does not exist is reported, not silently ignored. Every boolean also accepts "unknown", which ranks worst of all. Manifest schema.

Declare a ceiling

The ceiling is the part that catches what nobody anticipated:

ceilings:
  - agent: widgetworks-order-support-agent
    prohibited_resources:
      - customer-payment-methods
    prohibited_delegation_targets:
      - unapproved-agent

Any reachable capability that breaches it blocks the gate, even with every pattern rule suppressed. A ceiling breach cannot be suppressed. That is the first screenshot above. Capability ceiling.

Draft the manifest from the repository

uv run controlloop init . --dry-run

# controlloop.yaml
# ControlLoop security manifest -- generated by `controlloop init`.
...
schema_version: "1.1.0"

# TODO: owner
# ControlLoop cannot determine who is accountable for this
# repository's agents. Replace the placeholder below with an
# identifier for the responsible team or individual (for
# example: "platform-security" or "jane.doe").
owner: "TODO-owner"

trust_boundaries:
  # Discovered candidate agents (from repository discovery):
  #   - invoice-agent  (tools observed: Send an invoice)
  - name: "TODO-trust-boundary-name"

Every field is present even where ControlLoop could not determine a value; those are marked TODO with a comment saying what to supply. init never overwrites an existing manifest.

Build and validate the capability graph

uv run controlloop bom . -o acbom.json
uv run controlloop bom acbom.json --validate

✓ VALID  acbom.json is a valid ACBOM

For the reference agent that is 16 nodes and 8 edges, with a content digest the evidence artifact records. --format cyclonedx writes a CycloneDX 1.6 projection instead.

Machine-readable output

--format json on scan gives the complete report; on gate it is the evidence artifact itself, byte for byte. Every finding carries its own remediation and a stable documentation URL:

{
  "rule_id": "policy.undeclared-tools",
  "severity": "medium",
  "confidence": "high",
  "message": "Tool 'Look up an order' is reachable but not declared in controlloop.yaml.",
  "change_state": "not-compared",
  "remediation": "Declare this tool in controlloop.yaml (for example, an approvals entry naming its capability), or remove the agent's access to it if it was not intentional.",
  "documentation_url": "https://docs.controlloop.ai/reference/policy-rules#undeclared-tools"
}

--format sarif gives a SARIF 2.1.0 log for code-scanning tools, with the blast-radius statement as runs[0].properties.controlloop.blastRadius:

{
  "version": "2.1.0",
  "tool": "controlloop",
  "results": 3,
  "first_result": { "ruleId": "policy.undeclared-tools", "level": "warning" },
  "blastRadius": [
    "! WORST REACHABLE  financial-action: widgetworks-order-support-agent can invoke Issue a refund"
  ]
}

(Summarized from the real document: the version, the tool name, the result count, the first result, and the property.)

See the rules

uv run controlloop rules

• RULE PACK  controlloop-community 0.9.0 (24 rules)
• FINGERPRINT  169a7ceb80ffac4ccdb2ab95bc7491cb296f8736ded076f666a3020d3f1a342e

✓ policy.undeclared-tools  [active, optional]
    A tool an agent can reach, directly or transitively, that the repository's
    controlloop.yaml never declares.

✓ policy.destructive-or-financial-action  [active, mandatory]
    A change that gives an agent a destructive or financial action it did not
    previously have, against the approved baseline.
…

Run it as a GitHub Action

- uses: actions/checkout@v7.0.1
  with:
    fetch-depth: 0          # gate needs real history for a baseline
- uses: controlloop-security/controlloop@<ref>
  with:
    path: ${{ github.workspace }}

The action runs gate, fails the check on a blocking verdict, uploads SARIF to code scanning, writes a check summary, and keeps one pull-request comment with the authority statement current across pushes. GitHub Action reference.

Run it before every commit

repos:
  - repo: https://github.com/controlloop-security/controlloop
    rev: <pinned tag or commit SHA>
    hooks:
      - id: controlloop-gate

The hook runs the same gate before each commit and finishes within its documented time budget. Pre-commit hook.

Exit codes

Team

Want Team for your organization? Team is set up directly with each organization. Get in touch and we will get you started. Nothing in this section changes how the free CLI behaves: with team_enabled at its default of false, the CLI makes no network request.

The free CLI earns trust one repository at a time. Team turns the same gate into a shared control point across your private repositories:

• One policy across every repo. Your organization publishes a bundle; repositories may tighten it, never weaken it.
• An approved-components list. Every hook, MCP server, plugin, or permission rule a pull request adds is judged against what your organization approved, by package, version range, digest, and pin. A grant that is not on the list is Review required until someone approves it from the queue.
• Exceptions with an owner and an expiry. A build that passed yesterday fails today with the date, not a mystery.
• Every agent, hook, and MCP server in one inventory, with the sanction state of each.
• Signed evidence for any period, for the auditor and for your tooling.

Every grant across the fleet, which host holds it, and whether your organization approved it.

An unapproved MCP server waits here. Approve it pinned to the exact version that was reviewed, or by pattern, and the next gate passes.

The Team gate on a real repository, with the real output of a sanctioned grant, an unapproved one, and no published bundle, is in Your first Team gate. To set up Team for your organization, get in touch.

Documentation

Start with the documentation index.

• Example walkthrough: six steps with the real output of each
• CLI reference: every command and flag
• Terminal report format: how to read each line
• Policy rules: all 18, with real output
• Capability ceiling
• Manifest schema: controlloop.yaml
• Material capability taxonomy
• Authority statement and Agent authority limits
• Adapters: what each source contributes and where it stops
• Evidence artifact format
• GitHub Action and pre-commit hook
• Configuration: every key and its precedence
• Privacy, security, and limitations
• Your first Team gate and What a Team gate sends

Support and questions

Have a question, need help getting set up, or want to share how you use ControlLoop? Join us on Discord.

Community

• Discord for questions and help
• GitHub Issues for bugs and feature requests
• SECURITY.md for reporting a vulnerability
• CHANGELOG.md for what changed

Contributing

See CONTRIBUTING.md.

For maintainers

The lockfile is authoritative. Do not install unpinned project dependencies outside uv.

uv sync --locked --all-groups
uv run ruff format --check .
uv run ruff check .
uv run mypy
uv run pytest
uv build

uv run pytest takes roughly two to three minutes locally, including real PyInstaller builds and a wall-clock performance benchmark that is sensitive to machine load.

License

Apache License 2.0. See NOTICE for attributions.