LinkedIn automation as an MCP server over stdio. Drives your installed Chrome — no separate app, no browser download.
LinkedIn MCP has well-structured tool definitions with consistent naming (verb_noun pattern), comprehensive descriptions (avg 180+ chars), and complete input schemas using Zod. All 17 tools are explicitly registered with types and descriptions. However, output schemas are not documented in the source, LLMs cannot see what fields to expect from responses. Error handling is present but generic (McpToolError without recovery guidance). Parameter descriptions are solid but lack format constraints (e.g., URL patterns, character limits). Security is well-gated (LINKEDIN_ALLOW_MUTATIONS), but no tool annotations (readOnlyHint/destructiveHint) are present. Composition is clean, each tool has one responsibility, and naming clearly distinguishes similar operations (e.g., search_people vs search_jobs vs search_companies).
Apply to a job posting via Easy Apply. Supports a dry-run mode to preview questions without submitting, and an answers map to auto-fill screening questions. This is a WRITE action, gated by LINKEDIN_ALLOW_MUTATIONS.
Post a comment on a LinkedIn post. This is a WRITE action, gated by LINKEDIN_ALLOW_MUTATIONS.
Fetch the signed-in user's messaging conversations. Returns a list of conversation summaries (participant, last message, timestamp).
Fetch the signed-in user's home feed. Returns a list of posts with metadata (author, content, engagement counts, timestamps).
Fetch the signed-in user's notifications. Returns a list of notification items (type, actor, timestamp, action URL).
Open and scrape a LinkedIn member profile. Returns normalized profile data: name, headline, location, about, experience, education, skills, and connection count.
Output schemas not documented. LLMs cannot see what fields responses contain (e.g., what does linkedin_get_profile return? profile object structure?). Forces agents to guess field names for downstream operations.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot distinguish safe reads from destructive writes at a glance. linkedin_send_message and linkedin_apply_to_job should be marked destructiveHint=true; linkedin_get_profile should be readOnlyHint=true.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Return the current quota status: remaining applications today, reset time, and any active rate-limit windows.
Start the manual login flow. Opens the Playwright-controlled Chromium window (headed) on the LinkedIn login page, optionally pre-filling the given email/password, so the user can authenticate manually including 2FA/captcha. Waits until the session is established or times out, then persists the session.
Clear the stored LinkedIn session: deletes the saved storageState and clears the persistent browser profile cookies so the next action requires a fresh manual login.
React to a LinkedIn post (like, celebrate, support, love, insightful, funny). This is a WRITE action, gated by LINKEDIN_ALLOW_MUTATIONS.
Run a Companies search and return normalized company cards (name, industry, location, employee count, company URL).
Run a Jobs search and return normalized job postings (title, company, location, posted date, workplace type, salary if shown, job URL).
Run a People search and return normalized result cards (name, headline, location, profile URL, connection degree). Supports a free-text query plus common filters.
Send a connection request to a LinkedIn member. Optionally include a note (max 300 characters). This is a WRITE action, gated by LINKEDIN_ALLOW_MUTATIONS.
Send a direct message to a LinkedIn member. The recipient must be a 1st-degree connection or have an open profile. This is a WRITE action, gated by LINKEDIN_ALLOW_MUTATIONS.
Return the current driver/session status: lifecycle state, whether a LinkedIn session is authenticated, and whether the persisted session cookies are still valid. Useful for deciding whether a login is needed.
Update the SIGNED-IN user's OWN profile. Edits only the fields you provide (all optional; supply at least one): firstName, lastName, headline, location, about. Drives the "Edit intro" and "Edit about" modals and returns a per-field outcome (updated / unchanged / failed). This is a WRITE action, gated by LINKEDIN_ALLOW_MUTATIONS. Structured sections (experience, education, skills) are not yet supported.
Error handling lacks recovery guidance. McpToolError is thrown but responses do not suggest next steps (e.g., 'Profile not found. Try linkedin_search_people() first.'). Agents cannot self-correct.
Parameter format constraints missing. 'profileUrl' accepts 'Full profile URL or /in/ slug' but no regex or format validation is documented. 'note' in linkedin_send_connection says 'max 300 characters' but no minLength/maxLength in schema. LLMs cannot validate before calling.
Pagination not fully specified. linkedin_search_people, linkedin_search_jobs, linkedin_search_companies accept 'limit' but no offset/cursor or total_count in response. Large result sets may be truncated silently.