---
name: agenko
description: Wire Agenko's trust boundary into AI agents — init(), @check receipts, @approve gates, OpenTelemetry tracing, and self-hosting. Use when adding receipts or approval gates, looking up Agenko env vars, integrations, or API endpoints, or checking what Agenko actually ships versus backlog.
license: Apache-2.0
metadata:
  homepage: https://docs.agenko.app
  source: https://github.com/arovi-labs/agenko
  docs: https://docs.agenko.app/llms.txt
---

# Agenko

Deterministic trust boundary for AI agents — **receipt after** (`@check`), **gate before** (`@approve`). OTEL-native. Apache 2.0.

## When to use this skill

- Adding `init()`, `@check`, or `@approve` to an agent
- Returning a claimed-versus-actual receipt for a consequential action
- Pausing deletes, refunds, deploys, or sends until a human approves
- Looking up env vars, framework auto-instrumentation, or control-plane endpoints
- Answering "does Agenko ship X?" without trusting marketing copy

## Core rules

1. **Docs first** — fetch current docs before writing SDK code. Do not implement from memory.
2. **Shipped vs backlog** — only claim what the [Introduction status table](https://docs.agenko.app/introduction) says is shipped. `why()`, harness objects, and the eval harness are not shipped.
3. **Measurements only on receipts** — grounding scores are always `estimated: true`. No LLM-as-judge on a receipt.
4. **Symmetric SDKs** — Python and TypeScript export the same shape: `init`, `check`, `approve`, `get_client` / `getClient`.
5. **`init()` DX** — zero required args, reads env, never throws, uninitialized = no-op.

## Read the docs

Install this skill first, then fetch the page for the task:

```bash
npx skills add https://docs.agenko.app
```

| Task | Read first |
|---|---|
| First receipt | https://docs.agenko.app/quickstart |
| What good instrumentation looks like | https://docs.agenko.app/best-practices |
| Nouns (action, run, receipt, gate) | https://docs.agenko.app/core-concepts |
| Env vars | https://docs.agenko.app/environment |
| OpenAI / Anthropic / LangGraph | https://docs.agenko.app/integrations |
| Self-host | https://docs.agenko.app/self-hosting |
| REST + OTLP endpoints | https://docs.agenko.app/api-reference |
| Env + SDK quick reference | [reference.md](./reference.md) |

## Machine access

Fetch the full index:

```bash
curl -s https://docs.agenko.app/llms.txt
```

Section indexes: `llms-docs.txt`, `llms-sdk.txt`, `llms-integrations.txt`, `llms-boundary.txt`, `llms-operate.txt`, `llms-oss.txt`.

Every page is Markdown — append `.md` or send `Accept: text/markdown`:

```bash
curl -s "https://docs.agenko.app/quickstart.md"
curl -s "https://docs.agenko.app/check" -H "Accept: text/markdown"
```

Search when you do not know the page:

```bash
curl -sG "https://docs.agenko.app/api/search-docs" \
  --data-urlencode "query=How do I gate delete_account with @approve?"
```

## Minimal integration

```python
import agenko

agenko.init()

@agenko.check
def handle_ticket(ticket):
    return process(ticket)

@agenko.approve(when="irreversible")
def delete_account(user_id):
    db.delete_user(user_id)
```

```typescript
import agenko from "@agenko/sdk";

agenko.init();

const handleTicket = agenko.check(async (ticket) => process(ticket));

const deleteAccount = agenko.approve({ when: "irreversible" }, async (userId) => {
  await db.deleteUser(userId);
});
```

## Environment

```bash
export AGENKO_API_KEY=agk_...
export AGENKO_HOST=http://localhost:8000
```

Local stack: clone `arovi-labs/agenko`, `cp .env.example .env`, `make dev`. Console at `http://localhost:3000`.

Empty `AGENKO_API_KEY` is valid: the process runs and nothing crashes, but traces do not land in a project.

## Workflow

1. Read `/best-practices` before instrumenting.
2. Fetch the specific `.md` page for the API you need.
3. Fall back to `/api/search-docs` when the topic is unclear.
4. Verify against the repo specs: `agenko/spec/openapi/control-plane.yaml` and `agenko/spec/telemetry/semantic-conventions.yaml`.

## Anti-patterns

- Wrapping every helper in `@check` — wrap business actions, not plumbing.
- Putting an LLM judge on a receipt and calling it measured.
- Gating non-irreversible work and calling the queue a policy engine.
- Documenting `why()` or evals as shipped.
- Logging API keys or user PII in spans before redaction ships.
