linkedin-post

A wrapper around mcp-server-linkedin (Apache-2.0) that adds two tools: create_post and delete_post. MIT licensed.

It isn't a fork. The upstream package is a pinned dependency (==4.29.0) and runs unchanged. This project only registers extra tools next to upstream's.

Features

• All upstream tools (profiles, jobs, feed, messaging, ...) work as before.

• create_post(text, confirm_post) publishes a text post to your feed.

  • confirm_post: false is a dry run: the text is inserted and then the draft is discarded.

  • The post goes out with the share dialog's current audience (normally "Anyone"), and the result reports which one was used.

  • The result's status is posted, dry_run, outcome_unknown, invalid_text, editor_not_found or insert_failed. retry_safe: false means Post was clicked, so check the feed before calling again.

  • On a failure, trace_dir points at upstream's trace run (screenshot, URL, page title) under ~/.linkedin-mcp/trace-runs/. LINKEDIN_TRACE_MODE=off disables it.

• delete_post(post, confirm_delete) deletes one of your own posts. This can't be undone.

  • post is a post URL (including post_url from create_post) or a URN like urn:li:activity:<id>.

  • confirm_delete: false is a dry run that only checks the post exists and is yours.

  • It uses the post's "..." menu and LinkedIn's own confirmation. Someone else's post returns not_deletable.

Related MCP server: linkedin-mcp-server

Architecture

Hexagonal (ports and adapters):

src/linkedin_post/
  domain/
    entities/        PostText (validate + normalize), PostRef (URL/URN -> post page),
                     PublishResult / DeleteResult (status, posted/deleted, retry_safe)
    ports/primary/   Posting: create_post, delete_post
    ports/secondary/ SessionGate, PostComposer, PostRemover
    services/        PostingService: validate first, open the browser only for valid input
  adapters/
    primary/         FastMCP tools (create_post, delete_post)
    secondary/       upstream session gate, share-dialog composer, post remover, failure trace
  composition/       wiring, and the hook that adds our tools next to upstream's post tools
  common/            LinkedIn URLs
tests/               domain, service (with fake ports), MCP tools, remover in Chromium,
                     upstream hook

What comes from upstream instead of our own code:

• Session: signed-in Patchright browser (get_ready_extractor, get_or_create_browser).

• Text rules: normalize_message_text and _refused_character.

• Post URL/URN parsing: feed_payload.permalink_paths_from_payload.

• Failure diagnostics: debug_trace.record_page_trace and mark_trace_for_retention.

• Error handling: tool_errors.

• Text insertion: execCommand('insertText'), adapted from upstream's message_sender.py (Apache-2.0).

• No private APIs: no Voyager calls and no cookie replay, only the real web UI.

This relies on internal upstream names, so the version is pinned. Re-check composition/upstream.py and the adapters before bumping it; uv run pytest checks the hook.

Development

uv sync
uv run ruff check .          # every ruff rule (select = ["ALL"]), see pyproject.toml
uv run ruff format --check .
uv run pytest                # the remover test runs in a headless Chromium

With Nix, nix develop gives Python 3.12, uv and ruff, and runs uv sync for you.

CI (.github/workflows/ci.yml) runs the same on Python 3.12, 3.13 and 3.14. Ruff is pinned because ALL picks up new rules on every upgrade; bump it on purpose and fix what it finds.

Setup

Replace the original linkedin server. Only one process can own the browser profile, and a second copy would forward its tools to the first one without create_post.

claude mcp remove linkedin -s user
claude mcp add linkedin -s user -- uv run --directory ~/Projects/linkedin-post linkedin-post-mcp --no-auto-import

Then reconnect with /mcp.

Remove

Restore the original server:

claude mcp remove linkedin -s user
claude mcp add linkedin -s user -e UV_HTTP_TIMEOUT=300 -- uvx mcp-server-linkedin@latest --no-auto-import