MCP server

Since January 2026, Vera ships an Model Context Protocol server. Any MCP-capable agent framework — Claude-based agents, LangGraph, custom orchestrators — can delegate constraint checking and verified decision-making to Vera instead of letting a language model improvise it.

This inverts the usual integration question. You don't have to replace your agents with Vera; you give your agents a tool that makes their decisions provable.

Tools

Tool	Purpose
vera_reason	Full pipeline on a natural-language task; returns the recommended action with its trace.
vera_evaluate_plan	Pure solver check of a candidate plan; returns per-rule verdicts and the unsat core on failure.
vera_parse_rule	Translate a plain-English rule to structured form (preview only — activation stays explicit).
vera_get_rules	Read the active rulebook so the calling agent knows the constraint landscape.
vera_query_memory	Query entities, relationships and decision precedents from the knowledge graph.

Connecting

The server speaks standard MCP over stdio or HTTP. A typical client configuration:

{
  "mcpServers": {
    "vera": {
      "url": "https://<your-deployment>/mcp",
      "headers": { "Authorization": "Bearer vera_sk_..." }
    }
  }
}

Tool scopes follow the API key: a key without reason scope exposes only the read and evaluate tools.

The pattern that works

The calling agent handles conversation, gathering context, and orchestration. Whenever it is about to commit to an operational action, it calls vera_evaluate_plan first. If the plan is infeasible, the unsat core tells the agent exactly which constraint to fix — far better feedback than a generic failure, and the loop converges quickly.

Autonomy ceilings apply to MCP calls exactly as they do to the console and REST: a third-party agent cannot execute beyond the level your operators configured, and every call is logged in the same audit trail.