CTRLRun stops AI agents from taking wrong, restricted, or malicious actions in your workflows.
Every action is checked against your rules before it runs. Allowed actions go through.
Sensitive ones wait for a person. Forbidden ones are blocked.

Execution safety for AI agents. A Python library that sits between the decision to act and the call that acts.
A consequential action happens at most once, exactly as approved, and leaves a receipt.
When the outcome is unknown, CTRLRun says so instead of guessing.

Runs in production on a single file, or on Postgres across hosts. Apache-2.0.

pip install ctrlrun && ctrlrun demo

What it does

The model guesses. CTRLRun does not. The ticket says refund €500. The agent asks for €5,000, one extra zero. The tool is in its list, the arguments are well formed, and the model is completely confident. Nothing above the call disagrees, because nothing above the call is a check: a tool being callable is not permission to call it with those arguments.

CTRLRun is that check. It reads the arguments about to leave your process and answers what may happen to them: let it run, ask a human, or stop it cold. Four rules do the work, and each one is a test in this repository before it is a sentence here.

The third one is the half people forget. A correct €500 refund commits at the provider and the reply is lost coming back, so the agent retries. Retry libraries, agent frameworks and tool loops collapse this failed into I do not know what happened. CTRLRun keeps them apart: a lost reply is AMBIGUOUS, never FAILED, and a retry against an AMBIGUOUS effect is refused until a human, or a reconcile hook, says what happened.

$ ctrlrun demo
CTRLRun demo — five ways an agent action goes wrong, and what stops it.
Policy: refunds up to €1,000 are autonomous, up to €10,000 need a human, above that are denied.

1. Duplicate effect after a lost response

   refund €500  →  remote commits  →  response lost  →  effect: AMBIGUOUS
   agent retries the same refund
   ✗ BLOCKED — effect may already have committed; blind retry refused
   remote refund calls: 1
   only a human moves it on:  ctrlrun resolve refund:txn_1 --committed|--failed

2. Approval mutation

   agent proposes refund €2,000  →  human approves apr_0aa78e0380ba55d77a601dc782f57095 (bound to the action hash)
   agent executes refund €5,000  →
   ✗ BLOCKED — approved action ≠ requested action (mismatch)

3. Concurrent agents, same effect

   Agent A  reserve refund:txn_123  →  ACQUIRED  →  executes
   Agent B  reserve refund:txn_123  →
   ✗ BLOCKED — already reserved (in_progress)

4. Approval replay

   approval apr_dbc8bc6f06690cdf2e2c55a4e591ef3b used once  →  consumed
   same approval presented again                            →
   ✗ BLOCKED — single-use approval already consumed

5. Authority escalation

   human €100,000 delegable  →  finance agent €25,000  →  support agent €2,000
   support agent's grant: dlg_5f8d41938a3f29972d5489d676cd9edb
   support agent requests €50,000  →
   ✗ BLOCKED — outside the delegated grant (authority_constraint)
   remote refund calls: 0
   finance agent tries to delegate €50,000 under its own €25,000  →  refused (containment: constraints)
   support agent requests €1,500  →  authority permits it, and the policy asks a human (apr_f86eca24dd80206ab5189ccb1b62aa55)
   two axes, and an action needs both: the stricter of the pair wins

Receipts (8): .ctrlrun/demo/receipts.jsonl
Events:       .ctrlrun/demo/events.jsonl

Read them:    CTRLRUN_STATE=.ctrlrun/demo/state.db ctrlrun receipts

Where it stops. It does not detect prompt injection: it contains the consequence rather than reading the cause. It cannot promise exactly-once against a remote it does not control, it refuses to knowingly act twice, and it rolls nothing back. Receipts are chained, so an alteration is detected; a truncation at the end and a forged append are not, because the head that would catch them is a row in the same database, and closing that is what ctrlrun anchor is for. They are not signed: alteration is not authorship. The badge above means the declared guarantees pass in the setup they ran against, and it does not mean secure, safe, compliant, certified or audited: what the badge means · OWASP-AGENTIC-TOP10.md names the four entries this does not address.

If an agent only reads and answers, you do not need CTRLRun. The moment it can send, pay, refund, delete, deploy, grant, revoke, approve, submit, purchase or cancel, you do.

Use it in three steps

The animation above is this section, recorded against the real library: one policy file, two short programs, four commands, nothing staged.

1. Install it.

pip install ctrlrun

2. Write down what the agent may do. One file, ctrlrun.yaml. Amounts are integer minor units, so 50000 is €500. Both ends of every band are bound, because an upper bound alone lets a negative amount through, and a refund of a negative amount is a charge. Anything not listed is denied; there is no default-allow.

schema: ctrlrun.policy/v2

actions:
  stripe.refund:
    effect: "refund:{payment_id}"
    rules:
      - when: { amount_gte: 0, amount_lte: 50000 }
        decision: allow      # up to €500: the agent acts alone
      - when: { amount_gte: 0, amount_lte: 1000000 }
        decision: approve    # up to €10,000: a human decides
      - decision: deny       # above that: never

3. Wrap the call that has the consequence. The decorator names the action, the effect key names the consequence it has in the world, and the context names who is acting. stripe here is a stand-in that records calls instead of making them.

import sys

import ctrlrun


class FakeStripe:
    """Stands in for the provider: it records calls instead of making them."""

    def __init__(self) -> None:
        self.calls: list[tuple[str, int]] = []

    def refund(self, payment_id: str, amount: int) -> dict:
        self.calls.append((payment_id, amount))
        return {"id": f"re_{payment_id}", "amount": amount, "status": "succeeded"}


stripe = FakeStripe()


@ctrlrun.protect("stripe.refund", effect="refund:{payment_id}")
def refund(payment_id: str, amount: int) -> dict:
    return stripe.refund(payment_id, amount)


if __name__ == "__main__":
    with ctrlrun.context(agent="support-agent"):
        print("€500   ->", refund(payment_id="txn_1", amount=50_000)["status"])
        try:
            refund(payment_id="txn_2", amount=500_000)
        except ctrlrun.ApprovalRequired as pending:
            print("€5,000 -> a human decides:", pending.request_id)
            with open("request_id.txt", "w") as handle:
                handle.write(pending.request_id)
        else:
            sys.exit("the €5,000 refund ran without a human; the policy is not in force")
    print("calls that reached the provider:", len(stripe.calls))

The €500 refund runs on its own. The €5,000 one stops and names the request a human answers:

€500   -> succeeded
€5,000 -> a human decides: apr_63e80076f2cccfee52b17491a4b2e125
calls that reached the provider: 1

A human answers from the shell. The grant names the hash of the exact action the human read, and when it lapses. Ids, hashes and dates are generated per run; yours differ.

ctrlrun approve "$(cat request_id.txt)"

granted apr_63e80076f2cccfee52b17491a4b2e125 for sha256:22ec1c398e4b93d080b6cba61e5e11b0e21879552ac5dbf63c192d2b2e6af752
expires 2026-09-13T20:10:11.367Z

The agent presents it, then tries to spend it on something else. The first call is exactly what the human approved, and it runs. The second is the same approval with one digit changed, and it matches nothing:

import sys

import ctrlrun

from agent import refund, stripe

with open("request_id.txt") as handle:
    request_id = handle.read().strip()

with ctrlrun.context(agent="support-agent"), ctrlrun.with_approval(request_id):
    # Exactly what the human read: €5,000 on txn_2.
    print("€5,000 with the approval ->", refund(payment_id="txn_2", amount=500_000)["status"])

    # The same approval, one digit changed.
    try:
        refund(payment_id="txn_2", amount=900_000)
    except ctrlrun.ApprovalMismatch:
        print("€9,000 on that same approval -> refused")
    else:
        sys.exit("a mutated action ran on a human's approval; that is the bug this exists to stop")

print("calls that reached the provider:", len(stripe.calls), "(the €9,000 never left)")

€5,000 with the approval -> succeeded
€9,000 on that same approval -> refused
calls that reached the provider: 1 (the €9,000 never left)

Every attempt, refusals included, left a receipt, and ctrlrun receipts lists them. That is the whole integration: a policy file, a decorator, a context, and with_approval to present a grant. Money is the example, not the scope. A condition is <argument>_<op>, so the same policy language reads role_in: [reader, viewer] or replicas_lte: 10 as easily as amount_lte, and nine domains below have one policy each. Protect your first action walks the same path with every output explained · Policy YAML reference · Cookbook: refunds, deploys, IAM, deletions, email, MCP.

Three ways to use it

You probably do not need an adapter. @protect covers anything running in this process: a raw model call, a LangChain tool, a hand-rolled loop, a cron job. The gateway covers anything that reaches its tools over MCP, in any language.

It works with agents you can and can't modify. WhatsApp, Slack and Teams bots, ChatGPT, Cursor, Codex, OpenAI Agents: any AI agent you have. CTRLRun checks the action, not the agent, so if the agent acts through a tool server or an API you run, the action is checked, and the agent is not rebuilt, redeployed or told. Agents you can't modify says where the boundary goes for each kind.

An adapter exists for one reason: to route an approve decision through the framework's own interrupt, so a human answers where they already answer. There is never a second place to say yes. ctrlrun-langgraph gives prevention, because the resumption carries the arguments and core re-checks them against the hash. ctrlrun-openai-agents gives attribution, because that SDK records that a call was approved and not what its arguments were. None of the three is only for agents: a worker, a webhook handler and a scheduled job cannot tell a first attempt from a retry either.

How it works

Every protected call, whichever way it arrives, goes through the same seven steps. Only then does it reach your systems.

  normalize  →  decide  →  approve  →  reserve  →  execute  →  resolve  →  record

1. Normalize: one action, one id. The call becomes an Action: a name, canonical arguments (sorted keys, no floats), a resource, the principal. Its SHA-256 is the action hash.
2. Decide: allow, ask or block. Authority first (may this principal propose this at all, and within what bounds?), then policy (how much autonomy does this action get?). Unknown action, missing policy or missing principal is deny. Silence is never permission.
3. Approve: bound to this action. A human answers against the action hash. The approval is single-use, expires, and matches nothing but that exact action, so arguments changed after the answer void it and a person answers again. Name a preconditions= provider and the approval is also bound to the resource state it was granted against, rechecked strictly before the reservation: that narrows the window between the answer and the execution, from minutes of deliberation to milliseconds. It does not close it, because the recheck is a network call and cannot run inside the atomic write.
4. Reserve: claimed once. The effect key, refund:txn_1 or namespace:prod-eu:checkout, is taken in one atomic write. A second caller, in another process or on another host, is refused.
5. Execute: your code runs. Only NotExecuted, raised by you, means FAILED; every other exception and every timeout means AMBIGUOUS. Deciding which one you are looking at is the hard part, so ctrlrun.transport does it for you: urlopen, HTTPConnection and HTTPSConnection from stdlib urllib and http.client, which raise NotExecuted only where the connection they opened was handed no request byte. After one byte, every failure stays the exception it was, and the outcome is AMBIGUOUS. No setting widens that.
6. Resolve: unknown is not failed. An AMBIGUOUS effect keeps its key and refuses a retry until ctrlrun resolve, or a reconcile hook that asked the remote, says what happened. Nothing runs twice on a guess.
7. Record: a receipt either way. A portable JSON receipt: who, what, decision, approval, effect key, outcome, and the hash of the policy that decided it, chained to the receipt before it. Refusals get one too.

Who may ask, and how much. The policy decides the action and cannot see who is asking. Who may ask at all is a second axis, authority: every principal needs a grant, a delegation cannot widen one, and an action needs both axes, the stricter of the pair. Since 0.9 a grant can also carry a budget, a metric with a limit over a rolling window, consumed on reserve inside the same write, so a thousand refunds that each pass amount_lte cannot add up to more than the grant allows. An AMBIGUOUS effect holds its budget until it is resolved, because otherwise an agent that can manufacture ambiguity could manufacture authority. A budget bounds what the next reservation may do; it cannot recall an action already in flight. Authority has the whole model.

State lives in SQLite by default, a file with no server and no ops, and the reservation holds across processes rather than merely across threads. Point it at Postgres when more than one host writes: pip install "ctrlrun[postgres]", one URL, the same guarantees graded by the same suite. Prove it in your own setup with ctrlrun verify, which runs the kernel's own failure scenarios against your policy in a scratch store. It reaches no network: the only sockets it opens are to the store you named and to loopback listeners it bound itself, which is how it grades the transport classifier.

The same shape in nine domains

Nothing in CTRLRun knows what a refund is. An action is a name, canonical arguments, an effect key and a resource, and the three questions asked of it are the same whichever domain it came from: how much autonomy does this action get, did a human approve this exact action, and has this effect already happened. Two things carry your domain, and you write both.

• The effect key is the only domain knowledge in the system. It is the string that says two calls are the same real-world consequence: refund:{payment_id}, namespace:{cluster}:{name}, grant:{user_id}:{role}, prescription:{patient_id}:{drug}. Name it well and a retry cannot act twice; leave it out and there is nothing for at most once to be about.
• Conditions are arguments, not amounts. The language is <argument>_<op>, so the same operators read replicas_lte: 10, role_in: [reader, viewer] and to_domain_eq: acme.com as easily as amount_lte. A band is available to a domain that has never issued an invoice.

Read any row left to right and it is one rule wearing different nouns. The security row is the one to read twice: adding a deny rule to a firewall is autonomous and adding an allow rule is not, which no amount threshold would have told you. The policy is where your judgement about your domain gets written down; CTRLRun is what makes it hold.

Documentation

docs.ctrlrun.dev is the documentation: concepts, guides, a cookbook, the full reference.

Contributing

Issues and pull requests are welcome: CONTRIBUTING.md and CODE_OF_CONDUCT.md have the working agreement, and SECURITY.md is how to report a vulnerability. Every claim in this file has a test behind it, so a change to the prose usually means a change to the suite. CHANGELOG.md and https://docs.ctrlrun.dev/ROADMAP say where it is going. Releases carry PyPI provenance attestations from GitHub Actions.

License

Apache-2.0. The enforcement kernel is and will remain fully open source.