Marketing overhaul: one positioning — the governed control plane for AI agents
2026-08-27
LOCKSTEP TRACEABILITY MATRIX --- api_endpoints: ["none"] sdk_methods_updated: ["none"] mcp_tools_updated: ["none"] ---
What We Built
Phase 2 of the website + docs rework: a content-only overhaul of the marketing site (site/src/*/index.html) that keeps the existing dark Fraunces/Figtree design system untouched. Three competing taglines collapse into one: "the governed control plane for AI agents" — routing, identity, budget, and evidence as one platform. The home hero leads with the new line and keeps the "your AI agents need HR" metaphor as support, not headline. The one-liner propagates through every page's pageMeta, the footer tagline, both llms.txt copies, agents.json, and ai-plugin.json.
The near-invisible differentiators are now surfaced with real, specific copy: the self-describing capability graph (GET /v1/capabilities), decision rationale (GET /v1/explain — no competitor exposes per-request "why this model"), tamper-evident governance evidence (SOC 2 / GDPR / EU AI Act / NIST / ISO 42001 control mappings), fail-closed per-request data-protection constraints, the A2A mesh, and revocable capability grants. New sections land on home, features, and developers.
The stale /models page (4 months old, 31/8 counts, dead model names) is rewritten around the real current catalog pulled from src/router/provider-catalog-pricing.ts: 9 providers led by the Claude 5 family (Opus 5, Sonnet 5, Fable 5), plus GPT-5.6, Gemini 3.x, Grok 4.x, DeepSeek V4, Sonar, Kimi K3, and GLM-5.
Why It Matters
The public face had drifted months behind the product. Prospects saw 31 models when we serve 40; they saw "Thompson sampling picks per request" when the live selector is capability-matched; they never saw the features that actually won the competitive bench (reliability + decision rationale). One coherent story, accurate numbers, and the real differentiators up front.
How It Works
No number is hand-typed. scripts/generate-catalog-counts.ts already emits site/generated/public-counts.json; site/vite.config.ts injects {{counts.models|providers|mcpTools|discoveryLayers}} into every page's meta and body. This phase extended the generator's self-healing phrase rules to the MCP tool count in both llms.txt copies (site/public/llms.txt and the hand-maintained duplicate in src/api/server.ts), so 116 can no longer drift by hand.
Dead marketing-origin links (/v1/ hrefs on the MNI and developers pages) now point at api.brainstormrouter.com. sitemap.xml gains /security/, /status/, /onboarding/ and uses canonical /legal/ paths. ai-plugin.json's legal_info_url is fixed to /legal/terms. The leftover site/vercel.json (read by nothing) is annotated as a Cloudflare-Worker parity reference only.
The Numbers
- Positioning taglines: 3 → 1.
- Stale count literals in built
site/dist/: 0 (31,8 providers,107,
117, 7-layer all gone; remaining matches are binary screenshots / dollar figures).
cd site && npm run buildclean;generate-catalog-counts.ts --checkidempotent.
Competitive Edge
Portkey and OpenRouter surface routing but never tell you _why_ a model was chosen and ship no tamper-evident governance artifacts. This overhaul puts BR's per-request decision rationale, the queryable capability graph, and fail-closed data-protection — none of which the field exposes — front and center.
Lockstep Checklist
> Content/data + build-wiring only. No API route, SDK, or MCP schema changed.
- [x] API Routes: none changed.
- [x] TS SDK: none changed.
- [x] Python SDK: none changed.
- [x] MCP Schemas: none changed.
- [x] Master Record: no capability added; existing capabilities only surfaced.