scale4-mcp-server

Anti-sycophancy startup-idea assessment protocol distributed as an MCP server.

Forces evidence-grounded scoring across seven axes, kills sycophancy by construction (you cannot score without disconfirming evidence), and emits a risk-weighted roadmap as the closing artifact.

Install

{
  "mcpServers": {
    "scale4": {
      "command": "npx",
      "args": ["-y", "scale4-mcp-server"],
      "env": {
        "SCALE4_SESSION_SECRET": "<32+ char random string>",
        "SCALE4_API_KEY": "<optional — Pro/Team features and quota enforcement>"
      }
    }
  }
}

SCALE4_SESSION_SECRET is required. Generate one with:

node -e "console.log(require('crypto').randomBytes(32).toString('hex'))"

SCALE4_API_KEY is optional. Without it, all free-tier tools work and no server-side quota is enforced. With a free-tier key the 3-assessment/month limit is enforced; with a Pro/Team key gated tools are unlocked.

Tools

| Tool | Tier | Description | |------|------|-------------| | scale4.start | free | Open an assessment session. Returns protocol spec + session token. | | scale4.submit_signal | free | Submit one piece of evidence for one axis. Validates shape (URL HEAD-check, length, placeholder detection). Returns refreshed token. | | scale4.check_progress | free | Per-axis completeness summary — how many signals collected vs required. | | scale4.score | free (gated) | Gated: runs only when all axes have minimum required signals. Returns 70-point scorecard + verdict. | | scale4.generate_roadmap | free | Risk-weighted milestone list from scored axes. Requires score() first. | | scale4.export | free | Markdown bundle of assessment + evidence ledger + roadmap. PDF format requires Pro. | | scale4.deep_research_mode | Pro+ | Returns optional-signal spec for an axis (competitor pricing, LinkedIn lookup, search volume). | | scale4.stackforge_handoff | free (build verdict) | Translates passed assessment into a Stackforge compose_stack intent. Gated by build verdict. | | scale4.pitchforge_handoff | Pro+ | Translates assessment into a Pitchforge deck-generation seed. |

The 7-axis protocol

Axis              Required signals
─────────────     ─────────────────────────────────────────────────────────
problem           user_quote ×3, counterfactual_quote ×1
market            icp_definition ×1, named_prospect ×3, acv_estimate ×1
solutions         named_competitor ×5, dissatisfaction_quote ×3, pricing_row ×3
wedge             wedge_statement ×1, anti_wedge ×1, replication_estimate ×1
distribution      channel_choice ×1, cac_benchmark ×1, founder_channel_fit ×1
economics         cac ×1, ltv ×1, gross_margin ×1, breakeven_customers ×1
founder           domain_years ×1, unfair_advantage ×1, shipped_artifact ×1

Maximum: 70 points. Verdicts:

• build — total ≥ 56, no axis below 6
• iterate — total 42–55
• scope_cut — 3+ axes ≤ 5
• pivot — any axis ≤ 3
• kill — pivot conditions AND total < 42

State posture

The server holds zero state. Every tool call takes a signed session token (srv1.<payload>.<hmac>), verifies the HMAC, and returns an updated token. The agent stores the token. Losing the token means starting over.

Expiry: 30 days from start. Token size limit: 32 KB (~80 full signal entries).

Anti-sycophancy guarantee

Every axis has at least one required disconfirming signal:

• problem — requires a counterfactual quote (someone for whom the pain is NOT acute)
• solutions — requires naming 5 competitors (not zero or one)
• wedge — requires explaining why incumbents haven't done this (specific reason required)

The scoring gate refuses to run until evidence is in. The agent can't simulate validation — it has to do it.

Security notes

• Session tokens are HMAC-SHA256 signed; any tampering breaks verification
• Axis JSON files are ed25519 signed; the public key is pinned in the binary
• SCALE4_TRUST_INSECURE=1 bypasses axis signature verification in dev only (NODE_ENV !== 'production')
• URL HEAD-check validator fails closed (4xx/5xx → signal rejected)
• Scoring expression evaluator uses a Pratt parser — no eval(), no Function()