Integration

MCP Server

MCP Server

Voyager ships an MCP server that exposes 60+ LinkedIn tools over the Model Context Protocol. Any MCP-compatible client — Claude Code, Claude Desktop, Cursor, or your own agent — can call LinkedIn and Sales Navigator through natural tool use.

The MCP server is the primary agent surface for Voyager. REST is available, but if you’re building on top of an LLM, MCP is what you want.

Setup

The Voyager dashboard’s Connect MCP card gives you a ready-to-paste claude_desktop_config.json snippet — it includes your tenant API key and default user ID, so it’s zero-config after paste. See li.scaleabm.org/dashboard.html.

Manual install

Clone the repo, install deps, and point your MCP client at the stdio server:

$git clone https://github.com/algomo/linkedin-voyager
$cd linkedin-voyager && bun install

Then add to your client’s MCP config:

1{
2  "mcpServers": {
3    "voyager": {
4      "command": "bun",
5      "args": ["run", "src/mcp-server.ts"],
6      "cwd": "/path/to/linkedin-voyager",
7      "env": {
8        "VOYAGER_API_URL": "https://li.scaleabm.org",
9        "VOYAGER_API_KEY": "voy_your_api_key_here",
10        "VOYAGER_USER_ID": "your-user-id"
11      }
12    }
13  }
14}
Env varRequiredDescription
VOYAGER_API_KEYYesTenant API key (voy_ prefix) — create one in the dashboard
VOYAGER_API_URLNoAPI base URL (default: https://li.scaleabm.org)
VOYAGER_USER_IDNoDefault LinkedIn user ID. Multi-user tenants can override per tool call

Smoke test before wiring into your client:

$VOYAGER_API_KEY=voy_xxx bun src/mcp-server.ts

The server speaks JSON-RPC over stdio — no network ports are opened.

Tool Categories

The catalogue lives in src/services/mcp-tools.ts and is the authoritative list. High-level breakdown:

Profile (4)

get_my_profile, get_profile, get_profile_viewers, get_member_badges

Search (5)

search_people, search_companies, search_jobs, search_groups, search_content

Messaging (9)

get_conversations, get_conversation_messages, send_message, mark_conversation_read, delete_message, edit_message, react_to_message, forward_message, get_affiliated_mailboxes

Connections & network (6)

send_connection_request, get_received_connections, accept_connection, get_relationship_status, follow_profile, withdraw_connection

Engagement & posts (9)

like_post, comment_on_post, create_post, delete_comment, delete_post, get_notifications, get_post_reactions, get_post_comments, get_posts

Feed (2)

get_feed, get_post

Companies (4)

get_company, get_company_employees, find_employee_by_title, get_company_posts

Groups (2)

get_groups, get_group_members

Batch (2)

batch_read_profiles, batch_read_relationships

Capabilities & session (5)

list_account_capabilities, get_account_capability, get_session_capabilities, get_network_summary, get_session_activity

Profile editing (3)

edit_my_profile, get_my_followers, get_my_following

Webhooks (2)

list_webhooks, get_webhook_deliveries

Sales Navigator (7)

salesnav_search, salesnav_get_lead, salesnav_typeahead, salesnav_get_account, salesnav_get_account_employees, salesnav_save_lead, salesnav_search_companies

Sales Navigator tools require li_a and li_ep_auth_context cookies synced in your session. See the SalesNav guide for details.

Tool Annotations

The DESTRUCTIVE_TOOLS set in src/mcp-server.ts flags mutations with destructiveHint: true so MCP clients (Claude Code, Claude Desktop) can prompt before running them. Current list:

  • send_message
  • send_connection_request
  • accept_connection
  • follow_profile
  • withdraw_connection
  • like_post
  • comment_on_post
  • create_post
  • delete_post
  • delete_comment

Every other MCP tool carries readOnlyHint: true.

A few write-shaped tools (mark_conversation_read, edit_my_profile, salesnav_save_lead) are currently annotated as read-only because they’re not in DESTRUCTIVE_TOOLS. Treat them as mutations in your agent logic and gate them with your own confirmation — don’t rely on the MCP client prompting.

Destructive tools perform real actions on LinkedIn. There is no undo for sent messages, connection requests, or published posts. Your agent should still confirm with the human before calling these — MCP hints are a second line of defence, not the only one.

Architecture

The MCP server (src/mcp-server.ts) is a thin adapter over the REST API:

  1. Loads the MCP_TOOLS catalogue from src/services/mcp-tools.ts
  2. Registers each tool with the @modelcontextprotocol/sdk McpServer class
  3. Routes tool calls to the Voyager HTTP API using your VOYAGER_API_KEY
  4. Returns result JSON as the tool’s text content
AI Client (Claude Code, etc.)
    │
    │ stdin/stdout (JSON-RPC)
    ▼
MCP Server  ──▶  Voyager REST API  ──▶  LinkedIn

Example Session

With the MCP server configured, you can describe intent in natural language:

You: Find CTOs at fintech startups in Berlin
Claude: I'll search SalesNav for CTOs at fintech startups in Berlin.
[Tool call: salesnav_search]
  filters: CURRENT_TITLE="CTO", INDUSTRY="Financial Services",
           REGION="Berlin", COMPANY_HEADCOUNT="1-50"
Found 47 leads matching your criteria:
  1. Anna Mueller — CTO at PayFlow (Berlin)
  2. ...

The LLM picks tools, chains them, and formats results — you never write an API call by hand.