operant-mcp

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Read-only MCP server for the OPERANT AI operating-agent calibration benchmark.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "operant-mcp": {
      "url": "https://operant-mcp.saagar210.workers.dev/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (5)

• get_results read-only

  Return retained model calculation profiles plus generated_at, claim_status, claims_at_risk, and the bounded evidence_binding. These rows are not durable named-model performance claims. Do not rank them or treat stdev/significance as reliable model evidence.

  Hints: read-only yes · destructive not declared · idempotent not declared · open-world no

• compare_models read-only

  Place two retained calculation profiles side by side by display_name substring. Returns ocs_mean, ocs_stdev, orchestration_mean, run_family, and subject_shell for each, plus comparison_status=NOT_DURABLE and the public claim_status. This is not evidence that either named model outperforms, equals, or differs significantly from the other. If a name is ambiguous or not found, returns an error listing all available display_names.

  Parameters:

  • model_a (required) — string — Display name (or substring) of the first model
  • model_b (required) — string — Display name (or substring) of the second model

  Hints: read-only yes · destructive not declared · idempotent not declared · open-world no

• get_methodology read-only

  Return the benchmark design: the 4 axes, the OCS formula (Youden's J), the 5 decision labels (PROCEED, PROCEED_SANCTIONED, REFUSE, ESCALATE, REROUTE), scoring blocks, and a concise what-it-measures / what-it-doesn't summary. Sourced from the baked case-file metadata.

  Hints: read-only yes · destructive not declared · idempotent not declared · open-world no

• list_cases read-only

  Return case metadata (no full task prompts): pair_id/id, axis, tier, grounding, and side indicators (malign/benign for axes 1/2/4; null for axis 3). Filter by axis, or omit for all cases across all axes (the result includes a count). Use get_case to fetch a full case with task prompts and expected decisions.

  Parameters:

  • axis — string — Axis to filter by: refusal-calibration | sanctioned-path | orchestration | escalation-reroute

  Hints: read-only yes · destructive not declared · idempotent not declared · open-world no

• get_case read-only

  Return the full case for a given pair_id (axes 1/2/4) or id (axis 3): malign and benign task prompts, expected decisions, grounding rationale, and bypass patterns. Axis 3 cases are single (unmatched) and use an 'id' field instead of 'pair_id'. Use list_cases to browse available ids.

  Parameters:

  • pair_id (required) — string — The pair_id (axes 1/2/4) or id (axis 3) to retrieve
  • axis (required) — string — The axis this case belongs to

  Hints: read-only yes · destructive not declared · idempotent not declared · open-world no

Repository

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.