MCP server for interacting with LinkedIn API to manage posts, images, articles, and profile information
The LinkedIn MCP server has 10 tools with explicit registration and schemas visible in src/mcp_server.py. However, there are significant quality gaps that prevent a higher score. Most tools have reasonable naming (verb-based: get_, create_) and acceptable descriptions (avg ~80 chars). However, parameter descriptions are inconsistent, some parameters like 'distribution' and 'lifecycleState' lack detail about expected structure or valid values. Critically, the server lacks output schema documentation entirely; there is no specification of what fields these tools return or how downstream tools should chain results. No error handling guidance is provided (e.g., what to do if a post fails, how to recover from invalid URNs). Parameter constraints are not formalized as enums (e.g., 'visibility' and 'lifecycleState' should be enums, not free-form strings). The 'linkedinVersion' parameter appears in many tools but is never explained, is it mandatory? What are valid values? The overall structure suggests a competent API wrapper, but it does not meet production quality for agent composition.
Create a LinkedIn post with article metadata (link preview).
Create a LinkedIn post with a single uploaded image.
Create a LinkedIn post with multiple uploaded images.
Create a reshare of an existing LinkedIn post.
Create a text-only LinkedIn post using the Posts API.
Fetch current profile info (id, names, headline, summary, location, picture).
No output/response schemas documented. Tools return unstructured text or JSON blobs without field specification. LLMs cannot determine which fields exist, what types they are, or how to chain results to subsequent tools.
Parameter constraints not formalized. 'visibility', 'lifecycleState', and 'distribution' accept free-form strings instead of enums. LLMs will hallucinate invalid values (e.g., 'visibility': 'PRIVATE' instead of 'PUBLIC'). No format constraints on URN parameters (author, parent, owner, imageUrn).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 48 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Retrieve user information for the authenticated user via OAuth2 userinfo endpoint.
Retrieve the identity verification report for the authenticated user.
Register an image upload and return the image URN and upload URL.
Upload an image binary to the provided upload URL.
Parameter descriptions are vague or missing for complex types. 'distribution' is described as 'Distribution settings for the post' but what fields does it contain? 'lifecycleState' has no explanation of valid values or default behavior. 'linkedinVersion' appears in 8 tools with no explanation of purpose, format, or required values.
No error handling guidance. No documentation of what errors each tool can throw, what they mean, or how agents should recover. E.g., if 'create_text_post' fails with invalid 'author' URN, what should the LLM do? Call get_profile first? Retry? Error messages are surfaced to the client as generic JSON-RPC errors without context.
Tool composition broken. 'initialize_image_upload' returns 'imageUrn' and 'uploadUrl', then 'upload_image_binary' accepts 'uploadUrl' and 'filePath', then 'create_image_post' accepts 'imageUrn'. But there is no documented return schema from these tools, so LLMs cannot verify the field names match or understand what to pass next. The workflow is inferred from parameter names, not documented.
No idempotency guidance for write operations. 'create_text_post', 'create_reshare', 'create_image_post', and 'create_article_post' are destructive and create side effects (posts to LinkedIn). If an LLM retries due to ambiguous error, it will create duplicate posts. No confirmation step or dry-run mode offered.
Credentials and secrets may appear in tool parameters. The 'linkedinVersion' parameter (which may be a version string or token) is exposed. No documentation of whether credentials are server-side injected or expected from the agent. The initialize method accepts an 'access_token' at startup, but agent tools should never require manual credential passing.
Minimal parameter validation. Required parameters are declared in 'required' arrays, but there is no validation of URN format, URL format, or enum membership. If LLM passes 'author': 'invalid-urn', the tool will accept it and LinkedIn API will reject it with a cryptic error. Early validation with actionable error messages is missing.