oleg-vdv/kepil

**Accountability layer for AI agents.** Give every agent a passport, put every
action through one gate, and keep a log that cannot be rewritten afterwards.
> 53% of organisations have had an AI agent exceed its intended permissions.
> 48% of agents in production run with no monitoring at all. Only 22% treat an
> agent as an entity with its own identity.
> — Cloud Security Alliance and State of AI Agent Security, 2026
Kepil is what the other 78% are missing: identity, mandate, enforcement,
evidence — and the part nobody else does, **undo**.
Русская версия: README.ru.md
---
What it does
**Passport.** Every agent version gets an immutable card: who built it, who runs
it, what it does, what it will *never* do, its risk class, its autonomy class,
its limits, and when its risks are due for review. A new version is a new card;
the old one is kept forever.
**Mandate.** A machine-readable power of attorney for one job: allowed actions,
allowed systems, spending limits, a validity window, and which action types must
be confirmed by a human. Anything not explicitly allowed is refused.
**Gate.** The single point through which an agent touches the outside world.
Every action is checked against the mandate *before* a model is even called.
Fail-closed: any error inside the check means refusal, never a pass.
**Journal.** Append-only JSONL where every record carries the hash of the one
before it. Editing or deleting a record is detectable — by anyone, using an
independent implementation:
**Undo.** The journal is a graph of actions, and every profession declares its
compensating action. Kepil walks that graph backwards and stops honestly at the
first step that cannot be undone. Agent platforms record what happened; this one
puts it back.
**Confirmations on your phone.** Irreversible actions arrive in Telegram with
two buttons — approve or return — so being accountable does not mean sitting at
a laptop.
Use it from any MCP client
Kepil ships an MCP server, so an editor, an assistant or another agent can work
through it — and every action still passes the same gate into the same journal.
Seven tools: list professions, create an order, run a step, read order status,
see what is waiting for a human, verify the journal, read an agent passport.
**One tool is deliberately missing: confirmation.** If a model could approve an
irreversible action, the human would drop out of the chain and the whole design
would be pointless. The confirmation card goes to a person — in the panel or in
Telegram — and no MCP client can press it. A test enforces this.
Guard your existing automations
Kepil has a small JSON API, so an n8n workflow, a Make scenario or your own
script can ask permission before acting:
The answer is recorded in the journal, so later you can show on what grounds the
automation did — or did not do — something. For n8n there is a ready node:
n8n-nodes-kepil.
**The API stays off until you set a token** (panel → Settings, or
KEPIL_API_TOKEN). A panel bound to localhost is protected by the binding; a
programmatic interface is not, so it is disabled by default.
An agent here is never fully autonomous
AgentPassport refuses to be constructed with the autonomy class where a human
can no longer cancel a decision. That is a deliberate architectural limit rather
than a missing feature — see
ADR-0002. The gate enforces the
same rule regardless of what a profession definition claims.
Professions: behaviour as data, not code
An agent's job is a JSON description: ordered steps, boundaries, limits,
irreversible action patterns, rollback rules. Adding a new kind of work means
adding a file — or filling in a form in the panel. The dangerous parts stay in