Updates

Ran a full production-readiness pass: every surface a paying stranger touches got closed to a finished standard. The shipped shape:

• Instant-key signup works end to end. The advertised one-call API-key path (POST /v1/auth/signup) mints a real key instead of 404-ing; the magic-link path still works alongside it.
• The MCP surface tells the whole truth. The public manifest now advertises all 17 tools — including the four execute_* live-trade tools (propose / approve / list / read) that were previously omitted. Live trading needs a broker-connected trade-scoped key.
• No surface fakes success. chat_deep degradation, the India RBI repo-rate sourcing, and quota refunds are honest; demo mode is fenced off from real money; background tasks report real state.
• One true coverage story. Marketing counts (asset desks, external services, 17 MCP tools) now derive from a single backing constant, and every page tells the same story: crypto is the deep desk; US equities (with options) and India equities are newer desks in Beta; macro and prediction-market data are context layers.

Up next: the marketing-home rebuild on these canonical numbers, the pricing calculator, and per-claim source attribution inline in chat.

Wrapped Q3 2026 plan today. The big shape changes since the prior update:

• Two domains live. Crypto (deepest) and US equities (covers stocks + options + crypto). T1 fast agent, T3 supervisor, and the alpha engine are all domain-aware.
• MCP server live. 17 tools at api.marketintell.ai/mcp/sse — chat, chat_deep, alpha_signals, signal_performance, screener, portfolio_*, sessions_list, billing_status, health, register*. Listed on Smithery + Glama; PRs open in awesome-mcp-servers + Anthropic claude-skills.
• Public accuracy. Every signal carries a 24h outcome on a public URL. Live 30-day rolling win rate at /track-record; 1.5KB embeddable widget at /widgets/accuracy.js.
• Cost loop closed. Per-tenant monthly_cost_cap_cents, per-domain rollups, Free tier profitable at published Anthropic prices.
• Eval loop closed. Production sampler + Claude-as-judge + golden bench in CI gating regressions.
• Engineering hygiene. 167 test files, 5 green workflows (CI / Deploy / Bench / Bench-Refresh / Load Test), Postgres-backed multi-instance rate limiter, GDPR data export + self-service deletion.

Up next: deepening options + earnings coverage, tightening the T3 cost loop, rolling out an in-Telegram Stripe Payments path.

Q3 2026 phase 1 + 2 of 4 wrapped:

• T3 supervisor wiring. Multi-perspective queries now route through a parallel-specialist supervisor with synthesis. SSE phases land as specialist_started → specialist_completed → synthesis_started for traceable reasoning.
• Portfolio-aware chat.The brain auto-loads the caller's portfolio (with explicit IDOR guards). Web pill, Telegram /portfolio_chat toggle, per-key default setting.
• Per-signal outcome capture. Entry snapshot + 24h + 7d resolution wired up across every signal. The public renderer ships in a follow-up; outcome data already feeds /track-record.

Q3 plan kicked off. Foundation phase wrapped:

• Cost loop closed. Every LLM completion writes a cost_tracking row; monthly_cost_cap_cents on api_keys; pre-flight check in the quota middleware.
• Eval loop closed.5–10% production sampling judged by Claude on a 4-axis rubric. Golden bench in CI fails on > 5% regression.
• Brain unification. Single router decides between T1 fast agent, T2 deterministic CoT, and T3 supervisor by query intent. The three orphan stacks became four routed tiers (T4 = autonomous agents calling T1/T2/T3 on a cron).
• Mastra version pinned. One-shot bun update @mastra/* so the rest of Q3 builds on stable primitives.