MCP Server

Your AI writes the code. Ours optimizes it. They talk to each other.

Connect your coding agent to Winnow with the Model Context Protocol. Inspect evidence, author governed drafts, run supported evaluations, and prepare owner handoffs. All from your IDE, without treating discovery as authority.

Package availability is separate from application deployment. If npm returns 404 for @winnow/mcp-server, the public distribution is not live; use a source checkout or wait for the guarded package-release receipt before using the npx configuration.

Start Free Read the Docs

mcp.jsonjson

// .cursor/mcp.json or ~/.claude/settings.json
{
  "mcpServers": {
    "winnow": {
      "command": "npx",
      "args": ["-y", "@winnow/mcp-server"],
      "env": {
        "WINNOW_API_KEY": "evmcp_your_key_here"
      }
    }
  }
}

The Winnow workflow. One protocol.

Everything your coding agent needs to author prompts, manage experiments, flags, metrics, and optimization workflows.

Prompts

Version and compare prompts from your IDE. Create a draft, bind it to an experiment, then take a supported result to its governed promotion owner.

Experiments

Create A/B tests, check statistical results, and prepare pause, resume, or winner handoffs for the authenticated experiment owner. Production changes stay with their Winnow owner.

Feature Flags

Inspect targeting, author gate drafts, and prepare rollout plans. Live exposure remains controlled by key scope, policy, and the owning Winnow checkpoint.

Optimization Agent

The flagship feature. Ask the agent to analyze experiments and recommend next steps. It reasons about your data through your IDE's LLM.

Guardrails

Check guardrail status, draft safety-rule changes, and trigger supported checks. Consequential changes retain their configured owner and receipt.

Metrics

Browse the metric registry, verify data flow, and prepare metric definitions with the context needed for correct instrumentation.

Quick Actions

Generate SDK instrumentation code, run model comparisons, and prepare governed Ship handoffs. Compound workflows without bypassing approval.

Code is a first-class surface

One governed state, available from the product and your coding agent.

MCP exposes supported evidence, authoring, evaluation, handoff, optimization-review, and monitoring workflows. Your host discovers the surface offered by the connected server. Key scope, workspace, backend policy, and owner receipts still determine which action can execute.

Core is the default when WINNOW_MCP_PROFILE is unset or blank and exposes 36 lifecycle tools. The variable accepts core, agent-evals, or legacy. Agent Evals exposes 42 tools, while explicit legacy preserves the 123-tool compatibility escape hatch. Compact profiles pass the current release payload and cold-list latency gates; use standard discovery after every restart.

coding agent → @winnow/mcp-server

// inspect and prepare from the IDE
// core is default; set agent-evals or legacy explicitly
prompt_draft(action="create_version")
prompt_draft(action="configure") // bind draft
experiment_create(action="create")
→ mSPRT running · SRM clean
winnow_ask(action="experiment_second_opinion")
→ verdict + could_be_wrong_if + evidence ptrs
release_handoff("prepare_winner") // owner decides

Read-only resources give your coding assistant ambient context; ask_agent bridges it to the optimization agent. State written through an authorized operation uses the same backend state as the app. A recommendation or handoff is not a production mutation.

New

Prompt authoring from your editor

Tell your coding agent to draft a new prompt, wire it into an experiment arm, and prepare a governed promotion handoff when the evidence clears. The registry tracks every version with full diff history; the owning Winnow checkpoint controls live promotion.

Each tool accepts an optional workspace_id. If your org has one workspace, Winnow auto-picks it and confirms back. If multiple, you get a list of options so your agent can retry with the right scope.

prompt_draft.create_versionDraft a version with model, provider, temperature, metadata

prompt_draft.configureLink a version to an experiment arm without bumping the version

release_handoffPrepare evidence for the governed promotion owner

prompt_inspect.diffStructural diff between two versions

prompt_inspect.list / versionsBrowse the registry from chat

prompt_inspect.getFetch by version number or live/candidate tag

How It Works

Three steps from install to collaborative optimization.

Install in 30 seconds

Add the MCP server config to your IDE. One JSON block, one API key. No local dependencies.

Talk to your platform

Ask your coding agent to inspect evidence, author a supported draft, or prepare a governed handoff. It calls Winnow through MCP.

Let the agents collaborate

The optimization agent surfaces recommendations through your coding assistant. Open the authenticated Winnow owner to approve or execute consequential work.

Beyond CRUD

Other platforms ship MCP servers that expose their REST API. That's useful but basic. Winnow's MCP server includes ask_agent, a tool that builds a full context bundle from your experiment state, statistical results, and agent history, then lets your IDE's LLM reason about it. Your coding agent and our optimization agent work together.

A maintained tool catalog across the Winnow workflow

Prompt registry with versioning, promotion, and diff

Read-only resources for ambient context

Prompt templates for guided workflows

ask_agent bridges your coding assistant and our optimization agent

Works with Claude Code, Cursor, VS Code, and any MCP-compatible host

One config block. One governed workflow.

Install in your IDE and start talking to your experimentation platform.

Start Free