Post to social media from OpenClaw

OpenClaw's docs describe it as "an open-source AI assistant that runs on your own hardware and meets you in every chat app you already use," with a gateway process bridging channels such as Telegram, Slack, Discord, Signal and WhatsApp to the agent.

For social posting that means you can ask from your phone in a chat thread, "post this to Bluesky," and get the post URL back in the same thread. OpenClaw saves third-party MCP servers in its config and, per its docs, projects them into runtimes that expose configured MCP tools.

OpenClaw's configuration docs put MCP server definitions under mcp.servers in ~/.openclaw/openclaw.json. Remote entries take url, transport (streamable-http or sse) and headers, and the docs' example builds the header from an environment variable the same way. Use streamable-http; that's what our server speaks.

Or save it from the CLI. The docs list openclaw mcp set <name> <json>, and openclaw mcp probe <name> connects to the server and lists its tools without running an agent turn. If the probe lists eight tools, the setup works.

// ~/.openclaw/openclaw.json
{
  mcp: {
    servers: {
      socials: {
        url: "https://socialsforai.com/api/mcp",
        transport: "streamable-http",
        headers: {
          Authorization: "Bearer ${SOCIALS_API_KEY}",
        },
      },
    },
  },
}

# or from the CLI
openclaw mcp set socials '{"url":"https://socialsforai.com/api/mcp","transport":"streamable-http","headers":{"Authorization":"Bearer sk_..."}}'
openclaw mcp probe socials

If the runtime you use in OpenClaw doesn't load MCP tools, any agent that can make HTTP requests can use the same features over REST. Give it the key and point it at our docs in plain text at /llms.txt. The endpoints mirror the tools: GET /v1/accounts, POST /v1/connect-links, POST /v1/posts (with dry_run and Idempotency-Key), GET /v1/posts/:id. Full reference in the docs.

curl https://socialsforai.com/v1/posts \
  -H "Authorization: Bearer $SOCIALS_API_KEY" \
  -H "Idempotency-Key: bsky-2026-11-01-1" \
  -d '{"accounts":["acc_..."],"text":"...","dry_run":true}'

openclaw mcp probe socials is the quickest check of what the agent will get: it should list list_platforms, list_accounts, create_connect_link, create_post, get_post, list_posts, cancel_post and get_usage. OpenClaw's toolFilter can trim that list per server. The inputs each tool takes are documented once on the hub, in every tool and its inputs.

Send these from any channel OpenClaw is connected to.

• "Post this to Bluesky: <text>. Dry run first and send me the preview." Calls list_accounts with platform: "bluesky", then create_post with dry_run: true. After you reply yes, the same call without dry_run and with an idempotency_key, and it sends you the post URL from the target.
• "Connect my Bluesky account." Calls create_connect_link with platforms: ["bluesky"] and sends you the URL. Open it on your phone and use an app password.
• "Every Friday, summarize what I shipped and draft a post, but don't publish until I approve." Each run calls create_post with dry_run: true and waits for your reply. Pass scheduled_at if you want it held for a set time.

Chat changes the shape of a session. You might send the request from Telegram, wait an hour, and approve from Slack, so the agent should carry the dry-run body forward unchanged and send it with the same idempotency_key it picked for the draft. If the gateway restarts and the agent sends the publish twice, the second call returns the first post with nothing new published.

What comes back to the thread is the receipt for each account: published with a bsky.app URL, or failed with an error and its fix. A connect link sent to your phone stays valid for 7 days by default, so you can approve the account later from the same thread. For an app that connects accounts for other people, external_ref on the link tags every account it creates, and list_accounts can filter by it.

Chat makes approval natural: tell OpenClaw to always reply with the dry-run preview and wait for "yes" in the thread. OpenClaw's toolFilter can also hide tools per server; exclude: ["create_post"] gives you a read-only setup for status checks. Our own checks, including the daily caps and the duplicate block, apply whichever channel the request came from; see keeping a human in the loop.

These are the failures that come from OpenClaw's side. When a tool runs and returns an error, the body has a code and a fix; the hub's error code list covers each one.

• openclaw mcp probe socials fails: check that transport is streamable-http, not sse, and that SOCIALS_API_KEY is set for the gateway process if you used the variable form.
• The tools don't show up in a run even though the probe works: the runtime may not expose configured MCP tools. Use the REST fallback above.
• Status checks work but the agent says it can't post: a toolFilter on the server entry excludes create_post. Remove it from exclude if you want posting from that setup.