← MCP server and skill grades

linkedin-mcp-server

stickerdaniel/linkedin-mcp-server

C
Type
MCP server
Scanned
Nov 12, 2026
Version
b847cc6ed35f
Files read
561

This tool connects AI assistants to your LinkedIn account through a local browser so they can read profiles, jobs, feeds, and private messages, as well as send messages and connection requests. The biggest risk is that by default it imports your live LinkedIn login from your everyday browsers and contacts the developer's server on every startup to load remote tools and AI instructions, meaning a compromised server, unpinned plugin update, or malicious LinkedIn page could read private chats or act as you.

What it can reach

Browser
Controls a local Chromium or Chrome browser signed into your LinkedIn account to view profiles, companies, jobs, posts, and messages, and to send messages, manage connection requests, archive chats, and dismiss jobs.
Secrets
By default, accesses your system credential store (such as macOS Keychain, Windows DPAPI, or Linux Secret Service) to decrypt and import live LinkedIn login cookies from installed browsers like Chrome, Brave, Edge, and Arc, and can also use optional proxy credentials.
Files
Reads local browser cookie files during setup and saves decrypted LinkedIn cookies, browser profiles, downloaded browser binaries, logs, screenshots, and background service data in ~/.linkedin-mcp and ~/.linkedin-mcp-daemon (or a custom folder whose parent directory can be deleted on logout).
Network
Connects to LinkedIn (optionally through a proxy), contacts mcp.daniel.sticker.name on startup for remote tools, reaches PyPI for update checks and latest plugin downloads, downloads browser binaries from Playwright's CDN, and can open local ports for HTTP access or a Docker login viewer.
System
Starts and manages background browser, installer, crash-watcher, and shared-service processes on your computer, plus virtual display and VNC server processes when using the Docker login viewer.

Findings

1 high5 medium8 low
  1. HighUndocumented startup phone-home and dynamic remote injection of tools and agent instructions

    linkedin_mcp_server/remote_catalog.py:60

    On every normal server startup, before any tool is called, the server sends an HTTPS request to https://mcp.daniel.sticker.name/oss/tools reporting the OS version, CPU architecture, Python version, install method, and system locale. If that server returns an Ed25519-signed catalog within two seconds, remote_tools.install dynamically appends up to 1,000 characters to the MCP server instructions (injected into the AI agent's prompt) and registers arbitrary remote tools—even allowing them to claim readOnlyHint=true and openWorldHint=false so clients may auto-approve them. Invoking any such tool posts its arguments to https://mcp.daniel.sticker.name/oss/tools/call (bypassing PROXY_SERVER) and returns unsigned text responses up to 256 KB directly to the agent, creating a remote prompt-injection and data-exfiltration path if the remote service or key is misused or compromised. None of this behavior is mentioned in README.md, manifest.json, or .env.example.

    Fix: Disable remote catalog fetching by default (require explicit opt-in via LINKEDIN_MCP_REMOTE_TOOLS=on), document the endpoint and data shared in README.md and manifest.json, force openWorldHint=true and readOnlyHint=false on remote tools, and do not append remote text to mcp.instructions.

  2. MediumUpdate check injects instructions telling the AI agent to overwrite pinned MCP client configs

    linkedin_mcp_server/update_check.py:217

    On the first LinkedIn tool call, UpdateNoticeMiddleware polls PyPI in the background to check for newer releases. When the installed package is behind, it appends a text block directly into the tool result instructing the AI coding agent to check the user's MCP client configuration file and, if a version is pinned instead of @latest, to 'fix it and restart the client.' In coding agents with file-editing access (such as Claude Code, Cursor, or Codex), this steers the agent to silently weaken a user's deliberate version pin without user consent.

    Fix: Report available updates purely as informational text for the human user and remove instructions telling the agent to inspect or edit MCP client configuration files.

  3. MediumTool error messages instruct the agent to fetch and follow an unpinned skill file from GitHub

    linkedin_mcp_server/error_diagnostics.py:41

    When page-read, network, or authentication errors occur, build_issue_diagnostics writes local notes (including hostname and local profile paths) and appends PACKET_GUIDANCE to the error returned to the AI agent. That guidance instructs the agent to read the local diagnostic notes, fetch and follow https://github.com/stickerdaniel/linkedin-mcp-server/blob/main/.agents/skills/issue-packet/SKILL.md from the mutable main branch, and search GitHub issues to draft a bug report. Directing an agent inside routine tool errors to fetch and follow instructions from an unpinned remote branch introduces an indirect prompt-injection risk if the repository branch is ever compromised.

    Fix: Keep diagnostic file paths informational only and remove imperative instructions telling the agent to fetch external skill files or search GitHub issues on routine tool errors.

  4. MediumPlugin downloads and runs the unpinned latest PyPI package on startup

    plugins/linkedin-mcp-server/.mcp.json:6

    Although the Claude Code, Cursor, and Codex plugin manifests declare a specific version (4.29.0), the bundled MCP configuration launches uvx with mcp-server-linkedin@latest. Whenever the coding agent starts the server, uvx queries PyPI and downloads and executes the newest published package and freshly resolved transitive dependencies without lockfile hash pinning. The README notes that automatic updates are enabled, but does not warn that if the PyPI package or any unpinned dependency is ever compromised, plugin users will automatically run that code on startup with access to their LinkedIn session and local browser cookies.

    Fix: Pin the package argument in plugins/linkedin-mcp-server/.mcp.json to the exact plugin release version, such as mcp-server-linkedin==4.29.0, and bump it alongside plugin.json during releases.

  5. MediumPlugin MCP configuration downloads and runs unpinned @latest package at runtime

    plugins/linkedin-mcp-server/.mcp.json:6

    Although the Claude Code, Codex, and Cursor plugin manifests declare version 4.29.0, the bundled .mcp.json launches uvx with mcp-server-linkedin@latest. Every time the host starts the MCP server, uvx resolves and executes whatever release is newest on PyPI rather than the version installed with the plugin; while the README notes automatic updates as a feature, it does not warn that a compromised PyPI publisher account or bad release will automatically execute arbitrary code with access to your LinkedIn session and local browser cookies.

    Fix: Pin the package specifier in plugins/linkedin-mcp-server/.mcp.json to the exact release version matching plugin.json, such as mcp-server-linkedin==4.29.0.

  6. MediumBundled plugin runs unpinned @latest package from PyPI on startup

    plugins/linkedin-mcp-server/.mcp.json:6

    Although the Claude Code, Cursor, and Codex plugin manifests pin release version 4.29.0, the plugin MCP configuration launches uvx with mcp-server-linkedin@latest. Every startup can pull a newly published package version and resolve open-ended dependency ranges from PyPI without a lockfile or hash check, meaning a compromised upstream release or transitive dependency would automatically execute code on the user's machine with access to ~/.linkedin-mcp.

    Fix: Pin the package version in plugins/linkedin-mcp-server/.mcp.json to match the plugin version (for example, mcp-server-linkedin==4.29.0) instead of @latest.

  7. Lowget_inbox is marked readOnlyHint=True despite clicking conversations and marking unread messages as read

    linkedin_mcp_server/tools/messaging.py:43

    While get_conversation and search_conversations deliberately omit readOnlyHint=True because clicking sidebar rows in LinkedIn's UI marks unread messages as read, get_inbox still declares readOnlyHint=True even though it navigates to the compose sidebar and clicks up to 50 conversation rows to extract thread URLs. MCP clients use readOnlyHint=True to auto-approve tool calls without asking the user, so calling get_inbox can silently clear unread indicators and trigger read receipts across the user's inbox.

    Fix: Remove readOnlyHint=True (or set readOnlyHint=False) on get_inbox and document in its description that enumerating thread IDs clicks conversation rows and may mark unread threads as read.

  8. LowAgent plugins run unpinned mcp-server-linkedin@latest on every launch

    plugins/linkedin-mcp-server/.mcp.json:6

    Although the Claude Code, Cursor, and Codex plugin manifests declare a specific version (4.29.0), their shared .mcp.json launches uvx with mcp-server-linkedin@latest. Every time the coding agent starts, uvx resolves and executes whatever release is currently latest on PyPI without user review, exposing plugin installations immediately if a future PyPI release is compromised.

    Fix: Pin mcp-server-linkedin to the exact release version in plugins/linkedin-mcp-server/.mcp.json and update it via scripts/prepare_release.py during version bumps.

  9. LowBundled agent plugins run unpinned @latest package from PyPI at startup

    plugins/linkedin-mcp-server/.mcp.json:6

    Although the Claude Code, Cursor, and Codex plugin manifests declare version 4.29.0, their shared .mcp.json launches uvx with mcp-server-linkedin@latest. Whenever uvx refreshes its cache on startup, the plugin downloads and runs the newest release from PyPI without asking for confirmation, even though the README advises AI agents to get user confirmation before enabling automatic updates. If a future PyPI release were ever compromised, existing plugin installations would automatically execute that code with access to the user's LinkedIn session and local browser cookies.

    Fix: Pin mcp-server-linkedin to the exact plugin release version (for example, mcp-server-linkedin==4.29.0) in plugins/linkedin-mcp-server/.mcp.json and keep it synchronized in scripts/prepare_release.py.

  10. LowMaintainer PR verification skill runs untrusted pull request code on the host

    .agents/skills/3-verify-pr-fix/SKILL.md:99

    This repository maintainer skill fetches a contributor's pull request branch (pull/$PR/head), checks out the commit, and runs the server directly on the host machine during optional live comparison. Because its diff checklist only looks for locale strings, CSS selectors, section navigations, and missing tests—rather than malicious code or pyproject.toml hooks—running live verification on an untrusted PR could execute arbitrary code with access to the maintainer's ~/.linkedin-mcp session. Normal users who install the MCP server or plugin never run this skill.

    Fix: Require a security review of the full PR diff (including pyproject.toml, scripts, and imports) before checkout, and run candidate PR branches inside an isolated container without mounting the maintainer's real ~/.linkedin-mcp directory.

  11. LowMaintainer review-triage skill acts on public PR comments and pushes without confirmation

    .agents/skills/triage-reviews/SKILL.md:52

    This repository maintainer skill fetches all issue-level comments on a pull request—which any GitHub user can post on a public repository—and instructs the agent to fix findings, run tests, commit, and git push without warning that comments are untrusted input or requiring human approval. A stranger could post prompt injection disguised as review feedback to trick a maintainer's agent into modifying code and pushing to the branch. Normal users who install the MCP server or plugin never run this skill.

    Fix: Explicitly treat PR and issue comments as untrusted data, ignore comments from unverified outsiders unless approved by the user, and require explicit human confirmation before running tests, committing, or pushing.

  12. LowMaintainer review-triage skill ingests public PR comments and pushes code without confirmation

    .agents/skills/triage-reviews/SKILL.md:21

    This repository-local developer skill fetches all issue comments on a pull request—which any GitHub user can post on a public repository—without marking them as untrusted data, and then instructs the coding agent to edit code, run lint and test commands, commit, and git push without a human confirmation step. This file is not shipped in the installed package or plugin, so it only affects contributors who invoke /triage-reviews in a local checkout, where a prompt-injected PR comment could steer the agent into running and pushing unintended code changes.

    Fix: Treat pull request and issue comments as untrusted data and require explicit human confirmation before running tests on modified code, committing, or pushing.

  13. LowRepository PR verification skill executes untrusted fork PR code on the host

    .agents/skills/3-verify-pr-fix/SKILL.md:99

    This maintainer skill in the repository audits pull request diffs only for LinkedIn selector and locale rules before checking out pull/$PR/head and running the server locally with uv run in the optional live comparison step. If a contributor or maintainer uses this skill on an untrusted external PR and approves the live check, arbitrary Python or build-time code from the PR runs directly on the host with access to saved LinkedIn sessions and developer credentials. Normal MCP and plugin installations do not include or run this skill.

    Fix: Add an explicit security review step for all changed code, dependencies, and build configuration before checking out or executing code from external pull requests, or run untrusted PRs inside an isolated container without host credentials.

  14. LowRepository review-triage skill processes public PR comments and pushes code without confirmation

    .agents/skills/triage-reviews/SKILL.md:21

    This repository skill fetches all issue-level PR comments—which any GitHub account can post on a public repository—without filtering by trusted reviewer roles or warning about untrusted prompt input, then edits files, runs tests, commits, and pushes changes without human approval. If a maintainer invokes this skill on an open PR, a crafted comment from a stranger could steer the coding agent into making and pushing unintended code changes. Normal MCP and plugin installations do not include or run this skill.

    Fix: Restrict comment ingestion to trusted reviewers or collaborators, treat comment bodies explicitly as untrusted data, and require human confirmation before running tests, committing, or pushing.