Skip to content

Quickstart

This takes about five minutes. You run the gateway locally, give it a skill, and walk through one learning loop with curl. No agent framework or API key is needed yet.

1. Run the gateway

git clone https://github.com/notaaaronfr/myelinate.git
cd myelinate
uv sync
MYELINATE_ENV=dev uv run myelinate serve --port 8080

pip install myelinate
MYELINATE_ENV=dev myelinate serve --port 8080

docker run --rm -p 8080:8080 -e MYELINATE_ENV=dev ghcr.io/notaaaronfr/myelinate:0.3.0

MYELINATE_ENV=dev turns authentication off and puts everything in a tenant called dev. Use it only on your own machine. Without it, Myelinate refuses to start until you configure credentials (see Security and SSO).

Check it's up, then open the interactive API docs at http://localhost:8080/docs:

curl http://localhost:8080/healthz
# {"status":"ok"}

2. Give it a skill

A skill is a named procedure made of short rules. You can import any existing SKILL.md (Anthropic, Hermes or agentskills.io format) or write one:

curl -X POST http://localhost:8080/v1/skills/import -H 'content-type: application/json' -d '{
  "agent_id": "support-bot",
  "markdown": "---\nname: refund-request\ndescription: Handle a customer refund request\n---\n## Steps\n1. Look up the order and confirm it is within the 30-day refund window.\n2. If the item was damaged, ask for a photo before approving.\n## Pitfalls\n- Never promise a refund before the order lookup succeeds."
}'
# {"changeset_id":"cs_...","status":"applied"}

3. Recall before a run

Your agent calls recall with the task. Myelinate returns the relevant skills and memories as ready-to-use instructions for your system prompt:

curl -X POST http://localhost:8080/v1/recall -H 'content-type: application/json' -d '{
  "scope": {"agent_id": "support-bot", "user_id": "alice"},
  "query": "customer wants their money back for a broken blender"
}'

The response contains a recall_id and instructions like:

## Learned skills
Rules learned from previous runs. Follow them when they apply; if one is wrong for this situation, don't follow it.

### refund-request
Handle a customer refund request

#### Steps
- Look up the order and confirm it is within the 30-day refund window.
- If the item was damaged, ask for a photo before approving.
#### Pitfalls
- Never promise a refund before the order lookup succeeds.

The query never mentioned "refund", yet the skill was found: retrieval matches on meaning-bearing words, word forms and common abbreviations. Configure embeddings for fuller semantic matching.

4. Record after the run

Send what happened, with the recall_id, so Myelinate can credit the skills that were used:

curl -X POST http://localhost:8080/v1/episodes -H 'content-type: application/json' -d '{
  "scope": {"agent_id": "support-bot", "user_id": "alice"},
  "recall_id": "<recall_id from step 3>",
  "task": "Refund a broken blender",
  "messages": [
    {"role": "user", "content": "My blender arrived broken, I want a refund"},
    {"role": "assistant", "content": "Sorry! Could you send a photo of the damage?"}
  ],
  "outcome": "success"
}'
# {"episode_id":"ep_...","reflection_scheduled":false}

The outcome is the most important field for learning. Use success, failure or partial when you know it, and attach a verification object ({"passed": true, "details": "..."}) when tests or checks ran. Users can also rate a run afterwards:

curl -X POST http://localhost:8080/v1/feedback -H 'content-type: application/json' -d '{
  "episode_id": "<episode_id from above>", "score": 1.0, "comment": "Great handling"
}'

5. See what it learned

curl http://localhost:8080/v1/agents/support-bot/insights

This shows each skill's uses and success rate (with a credible interval), running canary experiments, rules that keep being marked harmful, the learning lift versus the control group, and learning cost.

6. Turn on learning

Until now the gateway only stored and served knowledge. To have it learn (reflect on runs, write and improve skills, reconcile memories), give it an Anthropic API key and restart:

echo "ANTHROPIC_API_KEY=sk-ant-..." >> .env
MYELINATE_ENV=dev uv run myelinate serve --port 8080

Record a few runs with real outcomes, especially failures with verification details. After every few episodes (or right after a failure) Myelinate reflects and proposes changes. Additive changes apply automatically; edits to existing rules wait for review in GET /v1/changesets. See Governance.

Next steps