Turn "we govern our AI" from a claim into auditor-grade evidence.
Keel is the operating layer for agentic AI on Azure, Microsoft Fabric, and Databricks. Every change a person or an agent makes carries a record of who did it, who approved it, and which controls fired, in the form that OSFI B-13, NI 52-109, and Law 25 reviewers ask to see.
We run our own controls on ourselves.
These figures are queried from the ledger that governs this company. The red one is the system flagging its own owner; the evidence page tells that story in full.
Live figures from the ledger, as of 2026-08-30 UTC. They move.
Every organization adopting AI agents faces the same two questions.
When the auditor asks what changed in March and who approved it, a folder of screenshots and a governance memo are not an answer. A memo says the controls exist. A ledger shows each one firing.
Can I let agents in without losing control?
You cannot safely put an agent on an ungoverned estate. Keel installs the golden path first: gated deploys, protected branches, drift detection, rollback. It is the foundation agents can operate on safely, and the one your team needed anyway.
Can I prove what they did?
Every actor, human and agent, writes to one ledger held in your own estate. Who acted, under which identity, who approved, what the controls found. You read the same ledger your agents write to, row by row.
Every row below is Keel governing its own estate.
Two code repositories, a Databricks workspace, and the GitHub organization itself. Every change, deployment, and control run lands in the same table, whether a person, a pipeline, or an agent did it. There are no customers in this ledger yet.
| When (UTC) | Who | Actor type | Where | What | Run outcome |
|---|---|---|---|---|---|
| 2026-08-12 01:27 | Claude Code | agent | GitHub: keelport/keel | repo:keel | running |
| 2026-08-29 23:59 | keel-agent-keelport[bot] | service | GitHub: keelport/website | keelport/website#20 | succeeded |
| 2026-08-29 23:48 | keel-ops | service | GitHub: keelport/keel | linked architecture decisions to their implementing commits | succeeded |
| 2026-08-29 23:45 | keel-agent-keelport[bot] (for steventchir) | agent | GitHub: keelport/keel | Recipe: Free Edition warehouses can refuse to wake on query | succeeded |
| 2026-08-29 23:45 | keel-agent-keelport[bot] (for steventchir) | agent | GitHub: keelport/website | The control register: steering and proof, side by side | succeeded |
| 2026-08-29 23:44 | keel-agent-keelport[bot] (for steventchir) | agent | GitHub: keelport/keel | The estate briefs the agent at session start (ADR-0034) | succeeded |
| 2026-08-29 22:57 | keel-agent-keelport[bot] | service | GitHub: keelport/website | keelport/website#19 | succeeded |
Queried from the Keel evidence ledger as of 2026-08-30 UTC. Work in flight first, then newest first.
Keel ships changes one opinionated way and keeps a record of everything that shipped.
One change, whichever actor makes it, moves left to right.
One manifest, whole estate
A single reviewed file parameterizes a client: platform, environments, identity, approvers, evidence store. Repos generate from templates and onboard in dependency order with one command.
Stood up on a new client in a day.Deploys that can't skip the gate
One merged commit flows dev to stage to prod; prod always pauses for a named human, who sees exactly what approving applies. No auto-approve path to prod exists. Not policy: structure.
A rollback is a governed deploy, not a bypass.The estate can't quietly change
Daily controls compare live branch protection, approval gates, and infrastructure against what the manifest declares. Hand-edits and out-of-band changes become recorded findings the next morning.
Drift is bounded by time, not by luck.One evidence base
Every control writes to one ledger: runs, events, alerts, rollbacks, tamper-evident and checkable after the fact. "Show me all unauthorized prod changes this quarter" is a query, not an archaeology project.
A control that isn't logged is a claim, not evidence.Agents with accountable humans
AI agents get their own identity, cryptographically separate from any person's, and every agent carries the name of the person accountable for it. Supervision becomes gates and queryable evidence instead of watching.
Your agents' compute bill grows with the work; your payroll doesn't.Governed to the exit
Create, govern, sync, roll back, retire. Offboarding is where orphaned access comes from, so leaving the fleet is as governed as joining it: one reviewed change retires a repo, and the record survives.
History is evidence: archived, never deleted.Half of Keel steers the action. The other half proves it.
Governance tooling usually picks one: guardrails that steer agents, or logs that describe them afterward. Steering without proof cannot pass an audit, and proof without steering just documents the damage. Keel is one substrate with both halves: steering controls, which act before the action, and proof controls, which record it. Every control below writes to the ledger you saw above.
Proof controls record what happened, as it happens.
The record an auditor asks for, and the one your own team needs the day something breaks. One row per shipped control.
Steering controls put the right context in front of the actor.
The newer half. It grows one merged control at a time; the rows marked in design have their decisions on record and their builds ahead of them.
Most vendors sell controls. Keel sells evidence that the controls fired.
See it on your stack.
A thirty-minute walkthrough of the live evidence, then what standing this up on your estate looks like.
Write to hello@keelport.caMail reaches the person who built Keel, not a queue. Say which stack your agents run on and the reply covers what the evidence layer looks like there.