PostLake

PostLake: the social operating system for AI agents

Give an agent one reliable way to run social. PostLake lets agents publish, schedule, monitor, discover, reply, moderate, and learn across X, LinkedIn, Instagram, TikTok, Facebook, Threads, Bluesky, YouTube, and Pinterest.

One connection gives an agent one normalised contract instead of nine platform SDKs, nine auth models, and nine incompatible result shapes. It can act with a hosted MCP server, a drop-in agent skill, or the REST API.

An agent can plan a campaign, validate each destination before it spends a

credit, publish or save a draft for approval, confirm asynchronous posts,

read conversations and comments, handle what needs a response, and use the

results to decide what to do next.

• Hosted MCP: https://api.postlake.dev/mcp
• Documentation: https://docs.postlake.dev
• Website: https://postlake.dev
• Machine-readable docs: https://postlake.dev/llms.txt
• MCP Registry: dev.postlake/social
• Grok Build: install postlake from the plugin marketplace, or add this repo as a marketplace source

Why agents use PostLake

One contract, every network

Post once or fan out to every connected network. Each target has the same state, URL, time, and error shape, so an agent handles the result once rather than writing platform-specific recovery logic.

Safe to let an agent act

validate_post provides a free dry run before publishing. Idempotency keys make retries safe. Owners can require approval on an OAuth connection, which forces every publish request from that agent into a draft and prevents it from approving its own work. confirm_post checks an asynchronous provider after it accepts a post, and every action reports the individual target result rather than hiding partial failure.

Not just a publishing endpoint

An agent can read a unified feed of notifications, direct messages, comments, and post performance. It can then reply, moderate, engage, research a topic, or inspect an account before it acts. Social management stays in the same workflow as publishing.

Human control where it belongs

OAuth makes the account owner approve a connection in the browser. get_connect_link lets an agent send them directly to that approval step. Owners can revoke an agent, manage channels, set agent limits, and require approval from PostLake. An OAuth agent cannot create an unrestricted account key to bypass those controls, and never needs a social-network password.

Start in minutes

Hosted MCP: Cursor, Claude, ChatGPT, Gemini, Copilot, and more

PostLake is a remote Streamable HTTP MCP server. Nothing is installed, run, or self-hosted. Add this server in your MCP client, then complete the PostLake OAuth approval once.

PostLake supports protocol revisions 2026-07-28, 2025-11-25, 2025-06-18, and 2025-03-26. Existing clients can retain their legacy handshake; newer clients use server/discover and per-request metadata. Both /mcp and the API-key compatibility endpoint /v1/mcp preserve agent identity, approval guardrails and publishing throttles. Tool annotations are conservative hints, not permissions. OAuth-connected or restricted agents cannot create unrestricted API keys to bypass owner controls.

{

"mcpServers": { "postlake": { "url": "https://api.postlake.dev/mcp" } } }

Cursor

Open Settings -> Tools & MCP -> New MCP server, add the configuration above, then select Connect for PostLake and approve the browser sign-in.

Claude Code

claude mcp add --transport http postlake https://api.postlake.dev/mcp

Grok Build

This repository is a Grok Build plugin. It ships the hosted MCP server and the agent skills. No local process, no shell hooks, no API key in chat.

Add this repo as a marketplace, then install postlake:

grok plugin marketplace add PostLake/postlake-mcp

grok plugin install postlake --trust

Inside Grok Build you can also browse /marketplace after the official xAI catalog lists it.

The MCP server is https://api.postlake.dev/mcp (Streamable HTTP). On first use, Grok opens PostLake OAuth in the browser. The account owner approves the agent once and can revoke it from https://app.postlake.dev/app/agents.

This plugin calls only:

• https://api.postlake.dev (MCP and REST)
• https://postlake.dev and https://docs.postlake.dev (docs the skills cite)
• https://app.postlake.dev (OAuth approval and dashboard)

openclaw/ in this repo is a separate OpenClaw package. Grok Build does not load it.

Other MCP clients

Add https://api.postlake.dev/mcp as a remote MCP server. The initial OAuth challenge opens the host's sign-in and approval flow. Full setup instructions: https://docs.postlake.dev/mcp

Drop-in skills: coding agents

Install PostLake's agent skills for Claude Code, Cursor, Codex, Windsurf, and other skills-compatible coding agents:

npx skills add postlake/postlake-mcp --all

Set POSTLAKE_API_KEY in the agent's runtime, then ask it to handle social work in plain language. The skills cover accounts, publishing, scheduling, media, inbox conversations, and analytics.

REST API: custom agent runtimes

Use https://api.postlake.dev/v1 from LangGraph, CrewAI, AutoGen, OpenAI tool calls, or any HTTPS client. The REST API and MCP server work against the same profiles, channels, posts, and safety rules.

What an agent can do

The MCP server currently provides 63 focused tools. They are designed around social outcomes rather than individual platform APIs.

set_my_name changes only the authenticated agent's display name. The account owner can set and lock that name in Agent Control. Renaming does not change the underlying API key or OAuth client identity.

Meta advertising and Partnership Ads tools are not currently available. Do not ask an agent to use branded-partner, ad-account, or Partnership Ads operations.

list_posts supports bounded cursor pagination and filters for status, agent, profile, network, dates, approval state, campaign, single or shared target composition, and a case-insensitive literal caption search with q. This lets an agent inspect publishing history without loading an entire account's posts into one tool response. The REST endpoint GET /v1/posts uses the same search and paging behavior.

edit_scheduled_posts_bulk changes up to 200 scheduled posts in one guarded operation. Preview first with dryRun: true, inspect the per-post before and after values, then apply with the returned snapshot. It can change captions, per-network captions, exact destination accounts, campaign and asset metadata, or shift scheduled UTC instants. If a selected post changes before apply, the whole edit is rejected. create_scheduled_calendar previews and creates up to 200 scheduled posts in one transaction, using explicit times or an elapsed-UTC cadence. Apply its preview snapshot with a unique idempotency key. The REST bulk cancellation endpoint also removes an exact selected batch atomically. Creating a new calendar and cancelling an old one remain separate operations.

Account-key-authenticated MCP callers can also use create_api_key to hand off to another trusted service. OAuth agents do not see or receive this tool because an account key would bypass their owner-set limits.

Cancellation, Social Removal And Upload Erasure

These are separate actions. cancel_post stops future publication and keeps uploaded files. delete_post requests supported network removals and reports every destination; Instagram and TikTok published posts must be removed in their own apps. Neither action erases uploads or refunds spent publishing credits.

Use delete_post with dryRun: true to preview destination support, then obtain owner confirmation and send the returned snapshot on apply. Accepted removals are remembered on partial results and not repeated on retry.

manage_media lists stored files with action: "list". Use action: "preview_delete" with a media ID to see affected posts and obtain a snapshot. Only after the owner confirms irreversible erasure should an agent apply action: "delete" with that snapshot. Drafts, approval requests and pending publications must stop using the file first. Source bytes and stored previews are removed; published social copies remain online. Storage failures remain retryable. More than 200 saved references require support-assisted erasure.

Uploaded assets retain the existing minimum of 90 days on every open account, with longer scheduled-reference protection and no current automatic expiry. Explicit owner erasure is an exception, not a routine expiry policy. Sources and previews use public URLs. Keep durable originals and never upload secrets. Full REST contracts and dashboard directions: https://docs.postlake.dev/media#deletion.

An agent workflow that does not break trust

1. Call whoami and list_social_accounts to understand the account, limits, connected channels, and current readiness.
2. If a human still needs to connect a channel, call get_connect_link and send them the short-lived approval URL. Do not ask them to hunt through a dashboard.
3. Call get_platform_capabilities and validate_post before creating a multi-network post. PostLake returns each target's exact constraint and fix.
4. Use create_post to publish, schedule, or save a draft for human review. If the owner enabled Require my approval, PostLake saves a draft even when the agent does not request one. Include an idempotency key so a retry cannot double-post.
5. If a platform is processing asynchronously, call confirm_post to obtain the provider-confirmed state without waiting for a public URL.
6. Use list_notifications, list_conversations, and read_comments to see what needs attention. Reply, moderate, or engage only where the platform supports it.
7. Use get_analytics and get_post_analytics to turn results into the next informed action, not a spreadsheet someone has to interpret later.

Direct messages

Facebook, Instagram, X, and Bluesky use one normalized conversation shape. Agents list threads with list_conversations, open one with read_conversation, finish it with mark_conversation_read, and reply with send_message. A message can include content for an attachment, shared media, or a provider payloa