What's new.

Docs grew 6x. Filled every section the sidebar promised: 4 platforms docs (ChatGPT, Claude Desktop, Claude Code, voice mode), 4 nutrition docs, 4 analytics docs, 3 reference docs (MCP profiles, OAuth scopes, troubleshooting). Plus a glossary and an account-and-auth guide.

Spanish parity. Every new doc shipped in EN and ES together — no more bilingual gap on the docs surface.

Drift guardrails. Three new CI checks: check:tool-count (catches stale tool counts in copy), check:recipe-demos (every recipe must have a demo block), check:docs-links (catches broken internal links). The 100-tool-vs-349 class of bug can’t return silently.

Docs index polished. Every configured section now renders at /docs, with a page count or a “Coming soon” badge — no more hiding the planned shape of docs.

/training/progress broken link fixed. Cited from quick-start but had no page behind it. Now exists in EN and ES.

Cookbook demos — Every recipe now shows what the conversation actually looks like in ChatGPT: the canonical prompt, the agent’s realistic reply, and the widget it renders. 50 demos shipped across training, nutrition, and analytics — 16 distinct widget kinds, faithful to the React bundles ChatGPT renders.

One new recipe — see-my-next-meal joins the nutrition cookbook for the “what should I eat next?” moment.

OG image pipeline unified — One renderer, one slug registry, build-time validation. Typos in ogSlug now hard-fail the build instead of silently shipping 404 social cards.

Landing voice refresh — Copy unified to first-person singular across the homepage; FAQ social handle updated to @_verxion_.

Infra — CI fix for the MCP tool-catalog dump: shared types are now built before the catalog dump consumes them on fresh checkouts.

Multi-device hero — iPhone, MacBook, and a browser window all in one composed scene. Each shows a different verxion conversation: logging mid-set, reviewing six months of progress, asking what’s on the diet today.

Tool-result cards — Workout plan, period summary, and diet plan rendered as structured cards. Same shape your AI sees when it calls verxion’s MCP tools.

Animated chat sequencer — The iPhone plays a real conversation in a loop: ask for today’s plan, log a Smith bench set with RIR and setup details, watch the AI track fatigue across sets.

Locale auto-detect — Visit verxion.ai from a Spanish browser and you land directly on the Spanish version. Manual override via the language switcher persists.

Page consolidation — From thirteen sections down to nine. Sharper signal, less filler.

Sign in with Apple or Google — No passwords, no verification emails. Your account is created on first sign-in.

Privacy by design — Encrypted in transit and at rest. One-click export. One-click delete. No third-party tracking, no ads, no resale. GDPR-compliant from day one.

AI client coverage — ChatGPT (via the verxion MCP App), Claude Desktop, Claude Code, Cursor, and any MCP-compatible client over HTTP.

290+ use cases live — Training (12 areas), nutrition (7), analysis (10). Every operation a serious athlete needs, exposed as MCP tools.

Free during beta — No credit card. No usage caps. Accounts created today keep their data and history when paid plans launch.

OAuth scope enforcement, end-to-end. Every tool call, on every profile, now passes through the same scope check. No “happens to work because the read endpoint forgot to gate”. Read, write, and destructive tiers are uniform across the 5 resource groups.

Route-scope coverage audit. Every API route maps to a known scope; any unmapped route fails the build. The map lives in route-scope-policy.ts and is auto-tested.

Cron service migrated. Replaced the Railway-managed cron with a custom Node.js scheduler running on the cron-worker service. Retries, execution history, Sentry alerting included.

Beta hardening. Cookies, CSP, CORS, error handling, and account deletion paths all audited and corrected. The platform is ready for public beta.

Monthly snapshots. On the 1st of each month, verxion generates a consolidated view of the previous month — total volume, sessions, PRs, kcal, macro adherence, body changes, and an AI-generated narrative of what shifted and why.

Period summaries. Ask across longer ranges (“how was Q1?”, “year-to-date training”) and verxion aggregates monthly snapshots into a single report with totals, averages, and the arc across months.

New analytics endpoints. GET /period-summary?from=YYYY-MM&to=YYYY-MM and GET /exercises/:id/progress?period=3m for direct access without the snapshot detour.

Cron-driven. Snapshot generation runs as a scheduled job — no manual trigger needed. You can still regenerate a specific month if you logged data late.

Code Mode shipped. Instead of exposing every operation as a separate MCP tool, verxion now defaults to a single execute_code tool. Your agent writes a short TypeScript snippet that runs in an isolated sandbox, with every helper guarded by the same OAuth scopes as a direct tool call.

Why it matters. ChatGPT caps custom apps at 100 tools. Code Mode replaces that wall with a single, expressive primitive — no more “this app has too many tools” errors, and round-trips collapse: one call can read a routine, start a session, and log the first set.

Code Mode UI variant. For ChatGPT specifically, a hybrid profile combines execute_code with widget tools so session summaries, macro pies, and contribution grids still render as visual cards inline.

Security unchanged. Every internal call inside the sandbox enforces the same OAuth scopes it would as a top-level tool call. Code Mode lowers the prompt footprint, not the security bar.