The right amount. The right payment. Your rules.
Give agents the authority to request payments, while you define how much they can spend and when approval is needed.
Explore
IARPA Solutions Marketplace · Awardable 2026Ship autonomy.Stay in control.
Let AI agents act for yourbusiness and customers,with limits you set andproof anyone can verify.Let AI agents act for your businessand customers, with limits you setand proof anyone can verify.
Works with your stack
More AI. More access. More consequences.
Agents now act across systems, data, budgets, payments, and workflows at machine speed. When an agent goes off course, every connection widens the blast radius.
Explore the AI Incident Registry
Your agents can already do more. Now you can let them.
Refund a customer. Update a record. Change access. Decide which actions your agents can take, when they need approval, and independently verify the authorization evidence afterward.
Give agents access without giving away control.
Let helpdesk agents request a specific group change for a support or contractor account. Okta reference.
ExploreLet agents work with data. Keep sensitive changes in check.
Let administrative agents request changes to permitted fields and the checked version. Synthetic records.
ExploreShip the revision you approved. To the target you intended.
Review the software revision before an agent requests deployment to a designated service.
ExploreYour rules. Their actions. Your control.
Set the boundaries. Keel enforces them.
Keelai-platform-prod
Production
Every decision is a permit.Every permit leaves evidence.
Zero Trust in action: every request needs an AI Permit before it runs. Allow, deny, throttle, or review—each outcome leaves evidence your team can inspect later.
AI Permit
Record: permit_demo_01
Date: Jun 5, 11:57 PM PDT
Allowed
All checks passed
Support agent summarized a case and opened a follow-up ticket.
Authorized Cost
$2.87
Actual
Policy
pol_voice_intake v7
pol_voice_intake
Executed route
gemini-1.5-pro
Actor
support-agent
Support
Usage
22,300 tokens
Decision record captured
evt_req_demo_01_01 · Hash chained
Recorded usage · Provider receipt verified
$ keel-verify export ai-ptrm-a66f7f81.json --reportAI Permit-to-Restart-Machine — Verification ReportEvidence: VERIFIEDFinding: Permit decision ALLOW. Permit: a66f7f81-e371-4512-94f7-36ab51311d75 Decision: allow Issued: 2026-08-21T20:26:06.875563+00:00 Expires: 2026-08-21T21:26:06.875563+00:00 Subject: df277bd2-2fbe-4fc0-b7a3-cab4b1c8a4b3Authorized action Action: cloud.machine.restart Scope: d1503f92-bb36-4bc0-9a13-6c2bcaedeeaaAuthenticity ✓ Evidence bundle structure and signature validAuthorization ✓ Permit decision signed by the issuing key (issuance-time) ✓ Human review transition signed by the issuing keyEvidence coverage Permit decision signature: verified Human review transition: verified Operator approval: not provided Counter-signature: not provided Co-signer key assertion: not provided Co-signature quorum: not provided Managed MCP mapping binding: not provided Governance interpretation: not provided Structural hold evidence: not provided Dispatch eligibility: not provided Dispatch: not provided Closure: not provided Revocation: not provided Timestamp: not providedHuman artifact Status: Authorized, not dispatched Presentation: cloud_machine_restart.r1 (keel.presentation_registry.v17)Summary Keel authorized df277bd2-2fbe-4fc0-b7a3-cab4b1c8a4b3 to cloud.machine.restart on dedicated demo machine commitment sha256:ab80413e31aec14baf53596075dd01236b8d28b6f40d941b62036a5a41875fbb. The Permit was valid from 2026-08-21T20:26:06.875563+00:00 until 2026-08-21T21:26:06.875563+00:00. No provider dispatch is represented.Verifier 3.30.0Trust mode: (not determined)Semantics profile: keel.pre_pinning_default.v0Verified at: [local run time]Input: sha256:6b2d38f9f4c1fd5cfa0deb943413b7808d727b67610b4cfcdbba2c0609dc99a8An auditor in 2035 should need the evidence file and the verifier, not our website, our API, or our continued existence.
Logs can be edited.
Evidence can be verified.
Every AI Permit is backed by signed evidence. It shows what was requested, which checks passed, and what evidence is missing. Anyone can verify it independently without a Keel account. If verification depends on taking our word for it, we’ve failed.
Not alteredThe report shows that the exported evidence still matches the signed manifest.
Trust mode visibleThe trust mode is visible near the top, so weak modes cannot look stronger than they are.
Independently verifiedThe report shows the permit, authorized action, verified checks, and missing evidence.
One control layer for your agents
However your agents arrive, Keel adds a provider-neutral authority layer to your existing stack. Integrate once through a single API. Keel applies your rules, ties approvals to the exact action, and records authorization evidence you can independently verify.
Build
You build the agent yourself.
Vercel AI SDK · OpenAI SDK · Anthropic SDK · agent harnesses · your own code
Compose
Off-the-shelf parts, shaped by your own logic.
APIs · MCP tools · webhooks · tool hooks
Buy
Complete agents, straight from a vendor.
wherever they expose an API or tool hook
Copilots · vendor voice agents · workflow agents · vertical SaaS agents
Keel
Agent control infrastructure
Your rules. Your approvals. Your evidence.
Models · Tools · Data · Access · Workflows · Payments · Deployments
Keel doesn't observe your AI.
It decides if the request runs.
Built for production systems where control is non-optional.
Fails closed
No silent passthrough.
If the policy can't be evaluated, the request doesn't run. Safety mechanisms that fail quietly aren't safety mechanisms.
Enforced before execution
The invoice isn't the alert.
Budget and policy enforced before the request reaches your provider.
Decision-level audit trail
Hand it to your auditor. Walk away.
Every decision recorded with policy version, outcome, and reason - at the moment it's made.
Verifiable without us.
Your auditor walks the chain themselves.
No account. No API key. Even if Keel disappears.
Don't take our word for it.
Ask your AI if Keel belongs in your stack.
Have the AI you already use independently evaluate Keel, where it fits, its limitations, and whether you actually need it.
Copies the question and opens the selected AI service in a new tab.We provide the question. Your AI provides the answer, including if the answer is no.
You already started building the control layer.We finished it. One integration.Audit-ready on day one.
One integration. Authorization, budget enforcement, audit trails, and multi-tenant isolation — governed before execution, not observed after.
