multivon-mcp

[Docs](https://docs.multivon.ai/mcp) · Website · PyPI · multivon-eval (engine) · Changelog

These 23 tools cover what an autonomous eval agent needs to do its job: discover its own capabilities (eval_discover), normalize traces from supported sources (eval_ingest_trace), and run evaluators against them, with calibration evidence specific to the tested task. We put the framework behind an MCP boundary because eval belongs in the agent's working loop, not behind a separate dashboard.

An MCP server that gives AI coding agents direct access to evaluation tools. Drop into Claude Desktop, Claude Code, Cursor, Cline, or any Model Context Protocol–compatible agent.

When the agent is helping you build an LLM product, it can:

• Score a RAG output for hallucination without you writing the scaffolding
• Generate an adversarial PDF on demand to test your document AI
• Run the full pdfhell mini-suite against a model and analyse the results
• Produce a self-verifying audit pack with a SHA-256 file manifest
• Discover the full evaluation capability catalog as JSON

No copy-paste, and no asking the agent to figure out the SDK calls from python -c "..." one-liners.

Current release: 0.4.0. Tested with MCP Python SDK 1.29.x, multivon-eval 0.18.0, and pdfhell 0.6.2. See the changelog.

Install

pip install "multivon-mcp==0.4.0"

The next release carries this compatibility bound itself. Installation pulls multivon-eval, pdfhell, and the MCP SDK. The provider SDKs (anthropic, openai, google-genai) come along too — bring your own API key in env.

Configure your agent

Claude Code

claude mcp add --transport stdio --scope user multivon -- multivon-mcp
claude mcp get multivon

Set provider keys in your shell or secure environment before starting Claude Code. To share the server configuration with a project instead, use --scope project; Claude Code writes .mcp.json and supports environment-variable expansion there. It does not read claude_desktop_config.json.

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "multivon": {
      "command": "multivon-mcp",
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-...",
        "OPENAI_API_KEY": "sk-proj-...",
        "GOOGLE_API_KEY": "AIza..."
      }
    }
  }
}

Restart Claude. The 23 tools become available; ask Claude "use multivon to evaluate this RAG output" and it figures out which tool to call.

Cursor

.cursor/mcp.json or via Settings → MCP:

{ "mcpServers": { "multivon": { "command": "multivon-mcp" } } }

Cline / OpenCode / any MCP-compatible agent

Same shape — point at the multivon-mcp console script.

Local dev / debugging

From a clone of this repo:

mcp dev multivon_mcp/server.py

From a pip install (the file lives in site-packages, so resolve it):

mcp dev "$(python -c 'import multivon_mcp.server as s; print(s.__file__)')"

Opens the MCP Inspector UI in your browser. You can call any tool by name, see the JSON schemas, and watch the requests/responses.

The 23 tools

Discovery & document AI

RAG generation & retrieval

Safety, compliance, fairness

Agent & multimodal

Agent traces. eval_tool_call_accuracy and the other agent-trace evaluators in multivon-eval (ToolArgumentAccuracy, ToolCallNecessity, TrajectoryEfficiency, AgentMemoryEval, PlanQuality, TaskCompletion, StepFaithfulness) take an agent_trace=[AgentStep(...)] plus expected_tool_calls=[...] on the case. Three-shape semantics matter: expected_tool_calls=None skips, [] asserts "no tools called", and [...] checks for the named calls. On repository main (shipping in the next release), the MCP tool supports the same trace mode: normalize trace JSON with eval_ingest_trace, then pass its agent_trace plus expected_tool_calls to eval_tool_call_accuracy. Set require_order=true when sequence matters or penalize_unexpected=true for a strict allow-list. See the multivon-eval agent integrations for the source-of-truth tracer code.

Flexible scoring

Agent workflows (new in 0.3.0)

Example session

User: I just shipped a RAG endpoint. Can you check it for hallucinations?

Claude: I'll use multivon to evaluate it.
        [calls eval_discover to see what's available]
        [calls eval_faithfulness with your input/context/output]

→ score: 0.667 (passed: False), threshold: 0.9
  reason: 2/3 claims grounded
    ✓ "annual renewal" — supported by context
    ✓ "30-day notice" — supported by context
    ✗ "automatic upgrade" — NOT in context

Claude: Your RAG hallucinated the "automatic upgrade" detail. The context
        doesn't mention upgrades. I'd add a Hallucination evaluator to your CI
        gate, threshold ≥0.85, and re-prompt with explicit "only use facts
        from context" instructions.

Release decisions from saved evidence

Use eval_acceptance_report(report_json_path, policy_json_path) to apply the same multivon.policy/v1 contract used by the CLI and CI action. It makes no model calls. The result includes decision (accept, reject, indeterminate), exit_code, policy_digest, measurements and findings. Malformed files or policies produce an MCP tool error, not an accepted result. See the acceptance policy guide.

Comparison is diagnostic: eval_compare_runs retains identity_verified and identity_issues. Missing or incompatible evidence suppresses the paired p-value. A large p-value does not prove equivalence, and an unchanged failing application still needs an absolute release contract.

Evaluator responses include status, measured and original metadata. For skipped or errored measurements, score and passed are null. Do not coerce them into an ordinary pass or fail. Missing agent_trace means unknown; provide agent_trace: [] only for an observed empty trajectory. Trace ingestion also requires an explicit steps list.

In single-call accuracy mode, supplied expected arguments must match exactly: extra keys, absent null-valued keys and changed JSON types fail. Omitting expected_arguments explicitly selects tool-name-only comparison. Trace mode checks tool names/order, not argument correctness or external side effects.

These response semantics change in 0.4.0. Update clients that assume passed and score are always measured values. The stdio integration test exercises real initialization, discovery, acceptance, missing evidence and tool errors.

Why these 23 tools (not all 44)

eval_discover returns the full 44-evaluator catalog, so the agent can always introspect everything. The 23 tools we expose directly are the ones agents actually call mid-edit:

• RAG generation checks (faithfulness, hallucination, relevance, answer_accuracy)
• RAG retrieval checks (context_precision, context_recall)
• Safety / fairness guardrails (toxicity, bias)
• Compliance (pii_detection, schema_compliance) — local-only, no API egress
• Flexible scoring (g_eval, custom_rubric) for user-defined rubrics
• Multimodal (vqa_faithfulness, document_grounding) for vision agents
• Agent traces (tool_call_accuracy)
• Document AI (pdfhell_run, pdfhell_make) — for any RAG-on-PDFs flow
• Audit pack — when procurement is involved
• Discover — meta-capability for planning
• Agent workflows (compare_runs, generate_cases, ingest_trace) — the loop that turns one-shot scoring into iterative improvement

The three new 0.3.0 tools exist because evals pay off as a loop: generate a starting suite from your own docs (eval_generate_cases), run your agent over it, score the trace (eval_ingest_trace → eval_*), make a fix, then verify the fix improved things vs. the baseline (eval_compare_runs). Agents need that whole loop callable from within a conversation, or they fall back to ad-hoc judgment.

Exposing all 44 evaluators as MCP tools would bloat the agent's context window and overwhelm tool-selection. If you need an evaluator that's not directly exposed, the agent can still use multivon-eval as a library — eval_discover returns the import paths.

Evaluate saved execution evidence

The multivon-eval development checkout includes an OTLP evidence bridge. Its example instruments an actual MCP stdio call with the official OpenTelemetry SDK, grades the retained trace, then passes the saved report and policy to eval_acceptance_report in published multivon-mcp 0.4.0. The complete synthetic fixture is accepted; missing capture-completeness evidence is indeterminate. No model API calls are required.

This is an opt-in library integration, not automatic server instrumentation or a new MCP tool. The bridge is not included in multivon-eval 0.18.0; see its guide for the checkout installation, native capture requirements and limitations.

The development Gymnasium outcome bridge also produces reports this server can evaluate. Three actual stdio calls verified accept, reject and indeterminate decisions from saved SQLite outcome evidence. The adapter checks persisted state and forbidden changes; it is a development preview, not part of the server's pinned multivon-eval 0.18.0 dependency.

The development media evidence bridge binds native Inspect media to exact content and explicit verdict references. Published MCP 0.4.0 reproduced a rejection from its saved six-case document report over stdio. This tool reads the report and policy; it does not fetch or decode media itself. The new media APIs require the core development checkout and are not added to this server's pinned dependency by this example.

Dependencies

Tested runtime bounds (from pyproject.toml):

• mcp[cli] >= 1.29, < 2 — official MCP Python SDK and Inspector. MCP 2.0 has a different server API and is intentionally excluded until this server migrates.
• multivon-eval >= 0.18.0, < 0.19 — the 44-evaluator engine, current report schema, and reasoning-judge fix.
• pdfhell >= 0.6.2, < 0.7 — the 17-family mini-v4 registry, corrected trap renderings, and current audit-pack schema.

These bounds ship in 0.4.0. Upgrade with pip install -U multivon-mcp.

All Apache 2.0.

MCP server vs Claude Code skills vs eval-action — which one do I use?

multivon-eval ships three agent-facing surfaces. They overlap on what they call (the same evaluator catalog) but differ on where the agent lives.

If you're building an LLM product and want the agent in your editor to score a RAG output without copy-pasting Python, use multivon-mcp. If you live in Claude Code and want the bootstrap → audit → explain loop wired up as native commands, use the bundled skills. For PR-time gating, use the GitHub Action. Most projects end up using more than one.

The Multivon ecosystem

Four public packages plus one closed early-access product, built around the same evaluation engine:

License

Apache 2.0.

Citing

@software{multivon_mcp,
  title  = {multivon-mcp: MCP server exposing multivon-eval + pdfhell as agent-callable tools},
  author = {Multivon},
  year   = {2026},
  url    = {https://github.com/multivon-ai/multivon-mcp},
}

The experimental vector world-model bridge produces the same saved report format. Actual stdio checks using published MCP 0.4.0 and core 0.18.0 reproduced accept/reject/indeterminate for measured versus missing CartPole forecast checks. The server ran outside the development checkout so the installed core wheel was used. These post-hoc checks establish transport and coverage behavior, not world-model quality or a release decision. Training, simulation and prediction remain in the upstream library workflow; this server adds no world-model runner or heavy learning dependencies.