Colophon · the stack
What runs the pipeline.
Every model, MCP server, external API, and data source the engine depends on. This page is derived from agents/*/AGENTS.md frontmatter — add a new MCP server to an agent and it shows up here automatically.
Orchestrator
AgentForge runs everything.
AgentForge is a long-lived HTTP service (localhost:8010) that loads agent and workflow YAML, schedules nodes, manages parallel execution, enforces budgets, and propagates verdicts. NLT publishes content to it via Pattern A — idempotent self-registration through scripts/agentforge_publish_nlt.py. The engine is a consumer; AgentForge is the platform. The multi-lane poc-ship DAG — parallel build crew, four QA gates, a demand-gated prototype lane, and its regen loops — is composed of AgentForge primitives, not bespoke orchestration.
Models
Claude for reasoning, Gemini for vision.
Opus 5 handles the heaviest judges and open-ended builds (POC Director, Frontend Builder, and the hardware visual agents — Product Renders and 3D Modeler). Sonnet 5 runs the structured-output agents — analysts, adversaries, copywriters, designers, and most QA judges. Haiku 4.5 runs the cheap Engineering QA gate. Gemini 2.5 Flash Image (Nano Banana) generates the renders; Gemini 2.5 Pro judges them cross-family so a different model family is checking Claude's output. pe-firm declares Claude Sonnet 5, but the pe-evaluate workflow's firm node overrides it to Claude Opus 5 at runtime for that step.
banker, compliance-baseline, incorporator, poc-builder, poc-director, product-3d, product-renders
cad-designer, domain-and-brand, general, gtm-launcher, gtm-operator, linear-intake-create, local-ops-gateway, market-pricing-researcher, mvp-builder, mvp-builder-gateway, mvp-director, mvp-smoke, number-auditor, pe-bull-reply, pe-buyer-simulator, pe-competitive-analyst, pe-creative-director, pe-devils-advocate, pe-feasibility-analyst, pe-financial-modeler, pe-firm (→ Claude Opus 5 at runtime), pe-market-researcher, pe-regulatory-analyst, poc-content-qa, poc-copywriter, poc-designer, poc-gtm-qa, poc-post-deploy-audit, poc-product-designer, poc-prototype-builder, poc-prototype-builder-gateway, poc-prototype-designer, poc-prototype-qa, poc-signal-planner, poc-visual-qa
poc-engineering-qa
MCP Servers
Memory and research, wired into every agent that needs them.
Each agent declares its MCP servers in frontmatter. AgentForge injects them at session start. The PE research analysts pull primary sources through firecrawl and exa; every agent shares pattern-tier learnings through NLT Memory.
Intake-issue creation. Used only by linear-intake-create to mirror FUND verdicts into the engineering tracker.
linear-intake-create
Semantic search for primary-source research. Surfaces Reddit threads, news, and comparable raises before the analyst writes a citation.
pe-competitive-analyst, pe-devils-advocate, pe-feasibility-analyst, pe-financial-modeler, pe-market-researcher, pe-regulatory-analyst
Web fetch for primary-source research. PE analysts use it to read competitor pricing pages, regulatory filings, and industry reports.
pe-competitive-analyst, pe-feasibility-analyst, pe-market-researcher, pe-regulatory-analyst
Cross-session memory. Postgres-backed, exposed via the NLT Quality Pipeline's BrainBridge. Tiered retention (architectural 180d, pattern 60d, procedural 30d, context 14d) and hive-scope for cross-agent pattern sharing.
pe-firm
Recall-gated, high-cost web search for PE market sizing. The research-library runner caches every query so repeat asks of the same idea reuse a snapshot instead of re-spending; capped at one live call per run on a cache miss.
pe-market-researcher
External APIs
Every service the pipeline touches.
Inferred from each agent's credentials and secrets declarations. Mercury and Stripe Atlas only appear post-FUND on the portfolio-create path; Cloudflare, Plausible, and GA4 wire in at gtm-launcher's landing-page publish step — none of the five are in the /pipeline idea-to-demo loop.
Image generation (Flash Image, branded as Nano Banana) and cross-family vision judging (Pro). Cross-family means a different model family judges Claude's output, avoiding same-family bias.
GEMINI_API_KEY domain-and-brand, product-3d, product-renders
Source control + intake webhooks + PR comments. Read-only PAT (long-lived) for repo clones; see docs/operations/github-app-auth-runbook.md.
GITHUB_TOKEN mvp-builder, poc-builder, product-3d, product-renders
POC site deployment (SSR + static CDN). Auto-deploys on PR merge; per-POC service with the slug as subdomain.
RENDER_API_KEY mvp-builder, poc-builder
DNS A and CNAME records for *.nltlabs.ai. Each POC's subdomain is provisioned and bound to its Render service.
GODADDY_API_KEYGODADDY_API_SECRET poc-builder
DNS, edge routing, and CDN for portfolio domains. Used by domain-and-brand for new LLCs and gtm-launcher for landing pages.
CLOUDFLARE_ACCOUNT_IDCLOUDFLARE_API_TOKEN domain-and-brand, gtm-launcher
Privacy-respecting analytics for shipped POCs. Configured by gtm-launcher at landing-page publish time; no cookies, no third-party tracking.
PLAUSIBLE_API_KEY gtm-launcher
Business banking. KYC application for the LLC after pe-firm issues a FUND verdict. HIGH risk — explicit operator approval per spawn.
MERCURY_API_KEYMERCURY_SANDBOX_MODE banker
Delaware LLC / C-Corp formation. The portfolio-create path's first step after FUND. HIGH risk — explicit operator approval per spawn. Vault: STRIPE_ATLAS_API_KEY (ADR-016).
STRIPE_ATLAS_API_KEY incorporator
Image-to-3D generation — derives a rotating .glb model from the hero product render (hardware POCs only).
MESHY_API_KEY product-3d
Traffic analytics on shipped landing pages, alongside Plausible. Wired by gtm-launcher at publish time.
GA4_API_KEY gtm-launcher
Data sources
How the research library actually gets its facts.
Every PE research pull declares which backend answered it, resolved through the _BACKENDS map at services/research_library/runner.py.
Licensed comparable-deal data consumed as cached snapshots (bigideasdb-snapshot method) — never a live call to the hosted MCP endpoint (agents/_includes/pe-licensed-demand.md forbids that).
Self-hosted scrape backend for pages that resist Firecrawl (JS-heavy or rate-limited sources).
Bulk tabular data pulled from a published CSV/export URL (e.g. government or open datasets).
Structured JSON/REST fetch against a documented API endpoint — no scraping or search involved.
Live web fetch (scrape/search/map/crawl) for primary-source research — competitor pricing pages, regulatory filings, industry reports.
PDF-to-text extraction for regulatory filings, 10-Ks, and other document-only primary sources.
Recall-gated live web search for market sizing. Cached before every call; capped at one live call per run on a cache miss.
Derived from agents/*/AGENTS.md via scripts/build_colophon.py. Re-run on every agent change so the public colophon stays in sync with what's actually declared at the engine layer — no hand-curated lists drift. See /pipeline for what the agents do together.