Build an AI social agent

Build an AI agent
that runs your socials.

Most agent demos die on platform integrations. PostLake is the hands: one publish call. One analytics shape, MCP if you want zero wrapper code. Your model still decides what to say, PostLake only executes and reports.

Free tier. No card · one API for every network · full MCP access

Before you start

Same setup the docs quickstart uses. Do this once:

1. Sign up at app.postlake.dev and verify your email (unlocks free credits).
2. On Channels, create a profile (e.g. my-brand) and connect at least one account. Bluesky is the fastest first channel: no app review. Instagram/TikTok/Facebook need each platform's review before API posting.
3. Account menu → API Keys → create a key (sk_live_…). It shows once; treat it like a password.

Two tools: publish and learn

The agent needs hands (publish) and feedback (analytics). Both match the docs. Same auth. Same profile. Same response discipline:

import os, uuid, requests

HEADERS = {"Authorization": f"Bearer {os.environ['POSTLAKE_API_KEY']}"}

def publish(text: str) -> dict:
    """Publish to every account on the profile. Always check targets[]."""
    return requests.post(
        "https://api.postlake.dev/v1/posts",
        headers={**HEADERS, "Idempotency-Key": str(uuid.uuid4())},
        json={"text": text, "profile": "my-brand"},
    ).json()

def performance(period: str = "30d") -> dict:
    """Same metric names on every network, agent-friendly."""
    return requests.get(
        f"https://api.postlake.dev/v1/analytics?period={period}",
        headers=HEADERS,
    ).json()

# Or skip wrappers: connect https://api.postlake.dev/mcp (OAuth).

Make it work

Tool-specific glue on top of the shared setup above:

1. Connect the accounts the agent may manage under one profile (Bluesky first for a fast smoke test).
2. Expose publish (POST /v1/posts) and performance (GET /v1/analytics), or use MCP tools for both.
3. Prompt the agent to treat partial as mixed results and to never claim success without reading targets.
4. Close the loop: after a few days, call analytics and let the model adjust format/network mix.
5. For calendars, add scheduledAt (UTC, or a local time plus timezone) so the agent can plan ahead without a second product.

Read the response (don't skip this)

You get one Post with an overall state and a targets[] array. One entry per account. Always check each target; partial success is normal.

{
  "id": "post_a1b2c3",
  "state": "partial",
  "targets": [
    { "platform": "bluesky",  "state": "published", "url": "https://bsky.app/…" },
    { "platform": "linkedin", "state": "failed",
      "error": { "type": "invalid_request", "message": "…", "retryable": false } }
  ]
}

Where the post goes

Same rules as the docs. Pick one addressing style:

See Publishing: where to post.

Do more (same API)

Pitfalls specific to this path

Common questions

Go deeper in the docs

These guides stay short on purpose. Canonical behaviour lives here:

Also: Media · Errors · MCP · Analytics

Related guides

All guides · Full docs · llms.txt · Markdown