MCP 服务器

@trigvale/mcp

Validates startup ideas with a deterministic scorecard, evidence brief, and verdict before code is written, integrating with MCP-aware build agents to avoid building dead-on-arrival products.

README

@trigvale/mcp

MCP server for Trigvale — validate startup ideas before your build agent scaffolds code.

This is the official Model Context Protocol server for Trigvale. It exposes one tool, validate_idea, to MCP-aware build agents (Cursor, Claude Code, Codex, Claude Desktop). The intended UX is straightforward: your agent calls validate_idea before it writes a single line of scaffolding. A green or amber verdict greenlights coding; a red verdict surfaces the assumption you should test first.

What `validate_idea` returns

10-dimension scorecard — Venture Readiness Score (VRS, 0–100), computed deterministically in code from the model's per-dimension scores against published anchors. The model never picks the VRS itself.
Evidence brief — every item tagged by sourceKind ∈ { observed, inferred, missing, ai, user-claim } so you know what's grounded vs assumed.
Verdict — kill, pivot, test, or build. Build is rare by construction.
Founder-fit calibration — per-dimension deltas applied from the user's declared skill graph (declared inputs only — never extrapolates from past failures).
1–3 archetype assignments — clusters the idea into known patterns (e.g. "vertical-ai-saas", "agentic-workflow") with structural cautions for the cluster.
Live evidence — real-time pull from Reddit, GitHub, Hacker News, Stack Overflow, and dev.to for the idea's dominant archetype, refreshed every 6 hours.

Full output schema and methodology: https://trigvale.com/methodology

Install

npx -y @trigvale/mcp

You'll need a Trigvale agent token (Integrate plan, $99/mo). Get one at https://trigvale.com/settings.

Configuration

Drop this into your MCP client's config (mcpServers block):

Cursor (`~/.cursor/mcp.json`)

{
  "mcpServers": {
    "trigvale": {
      "command": "npx",
      "args": ["-y", "@trigvale/mcp"],
      "env": {
        "TRIGVALE_AGENT_TOKEN": "tvk_xxx"
      }
    }
  }
}

Claude Desktop (`~/Library/Application Support/Claude/claude_desktop_config.json` on macOS)

Same shape — replace tvk_xxx with your token from /settings.

Claude Code

Add the snippet to .mcp.json at the repo root, or use claude mcp add to register it interactively.

Self-hosted / dev environments

Override the API base via env:

TRIGVALE_API_BASE_URL=https://api-dev.trigvale.com npx -y @trigvale/mcp

Default base is https://api.trigvale.com (production).

Example usage in an agent prompt

"Before you scaffold this project, call validate_idea with the user's pitch. If the verdict is kill or pivot, surface the weakest assumption and ask the user whether to proceed anyway. If test, suggest the validation sprint Trigvale recommends before any code."

Why a separate validation step?

Building has gotten cheap. The bottleneck is no longer execution — it's deciding what to execute. Most ideas should be killed, pivoted, or tested before code is written. validate_idea makes that decision an explicit, repeatable step in your agent flow, with a deterministic verdict and source-tagged evidence you can show the user.

Pricing & plans

Plan	Monthly	Includes
Free	$0	1 verdict / month, lite preview
Starter	$12	10 verdicts / month, full brief
Validate	$39	Live evidence pipeline + sharpening + sprint generator
Integrate	$99	Everything above + agent tokens for this MCP server

The agent token required by this MCP is gated to the Integrate plan. Per-call entitlement is re-checked server-side on every /agent/v1/evaluate call — downgrades take effect immediately.

Full pricing: https://trigvale.com/pricing

Source

This is a read-only mirror of the MCP server source from Trigvale's private monorepo (the rest of which contains the proprietary rubric, infrastructure, and billing logic). Releases are cut on the private side and synced here. PRs are welcome — please open issues at https://trigvale.com/contact instead of GitHub Issues so we route them through the support workflow.