Skip to main content
The MCP server exposes Seyn’s knowledge to external AI agents via the Model Context Protocol, the open standard supported by Claude, Claude Code, Cursor, and a growing list of agent runtimes. The pitch in one line: your agents stop guessing how your business works, because they can look it up.

​
The tools

ToolWhat it returns
get_processesAll processes in the org: names, descriptions, step and rule counts. The orientation call.
get_process_rulesRules for a process, optionally filtered by step: descriptions, confidence, frequency, review status.
query_knowledgeNatural-language search through the full hybrid query pipeline; supports the same strategies as the API.
get_rule_provenanceThe complete audit chain for a rule: inference log, events, raw records. Your agent can show receipts too.
get_related_rulesKnowledge-graph neighbours of a rule, with relation types.
get_entity_timelineChronological events for an entity: what happened to this deal, in order, with before/after state.
get_entity_graphMulti-hop entity neighbourhood: the people, tools, and entities connected to a starting point, with the assertions that link them.
A typical agent flow: get_processes to orient, query_knowledge for the user’s question, get_rule_provenance when the user asks “why?”.

​
Transport, auth, audit

  • Transport: stdio (JSON-RPC 2.0). The server runs as a local process speaking to your MCP client.
  • Auth: a Seyn API key (SEYN_MCP_API_KEY environment variable). The key binds the session to exactly one organisation; every tool result is scoped to it, with no cross-tenant path. Same model as the public API.
  • Audit: every tool invocation is logged with operation, latency, and outcome. An agent reading your knowledge base leaves the same trail a human would.
  • Rate limiting: per-key sliding window; over-limit calls return a structured error with retry guidance.

​
Setup

{
  "mcpServers": {
    "seyn": {
      "command": "npx",
      "args": ["@seyn/mcp-server"],
      "env": { "SEYN_MCP_API_KEY": "sk_live_..." }
    }
  }
}
Like the SDK, the MCP server package is in private alpha: distribution is currently by tarball from the team rather than npm. Request access; configuration is identical once installed.

​
Error shape

Tools return the platform’s standard error format (NOT_FOUND, VALIDATION_ERROR, RATE_LIMITED, INTERNAL_ERROR) as structured tool results rather than protocol failures, so agents can branch on them.

​
Common mistakes

SymptomCauseFix
Every tool returns empty resultsThe API key belongs to an org with no active knowledge libraryVerify the key’s org in the dashboard; run extraction if needed
Agent answers without consulting SeynThe agent’s system prompt doesn’t tell it when to use the toolsInstruct it explicitly: “for questions about processes, query seyn before answering”
RATE_LIMITED during agent loopsAgents can burn a sliding window fast in retry loopsHonor the retry guidance in the error; add backoff to the agent harness

Authentication

API keys: creation, rotation, revocation.

Query

What query_knowledge runs under the hood.