Changelog

Direct persona conversations and a clearer docs front door — 2026-07-25

Talk now has a dedicated person-first conversation modal. Pick one named workspace audience for a direct exchange or compare two perspectives side by side with one shared question.

New — One-person and two-person comparison modes with separate transcripts
New — Named selectors, synthetic and grounding labels, bounded per-lane history, and total credit estimate before every question
Improved — Talk is a primary app destination rather than an advanced-only surface
Improved — Docs front door now starts with the category boundary, first result, and role-based paths
New — Dedicated completed-study Ingest page covering source-quoted dossiers, stated versus inferred claims, and current storage and deletion limits
Fixed — Beginner guide no longer ends in an API console, and synthetic responses are no longer described as human respondent language
Fixed — Public docs now expose working robots.txt and sitemap.xml

Interactive guides and evidence-gated SOW execution — 2026-07-25

Guides now turn bracket placeholders into real inputs with a live prompt and one-click copy. Embedded API consoles derive their controls from the live OpenAPI schema, show the full request object before anything runs, call Mavera directly, and render the full response object.

New — Mavera Guide in every signed-in app screen: workspace-aware, docs-grounded, read-only customer-success guidance with safe page actions and a full-Mave handoff
New — Interactive prompt builders across the operator guide and prompt library, with live bracket substitution and copy
New — Connected guides: authenticated workspace handoff, audience selectors, and dedicated docs-console API-key generation or rotation
New — Embedded app tour with full-screen, sidebar-only, input-bar, and Mave-home screenshots
New — Embedded API consoles with typed path, query, and body controls, API-key auth, request JSON, real execution, and response JSON
New — docs2.mavera.io custom domain with automatic signed-in workspace and audience bootstrap
New — API-key metadata selector with secure per-tab paste plus dedicated docs-key generation or rotation
New — cURL, Python, Node.js, JavaScript, Go, and PHP examples generated from the live request
New — Page prompt handoff to clipboard, ChatGPT, Claude, Cursor, and Lovable
New — Contextual product simulations across all 41 operator guides, with a working Mavera shell, sidebar, inputs, selectors, progress, reset, and required interactions
New — Safe admin walkthroughs for team invites, spend controls, Stripe's test card, and one-time API-key creation without production writes
New — Naming SOW compiler with a 26-item requirement ledger and 15-module evidence graph
New — Durable SOW lifecycle endpoints for compile, create, inputs, approval, start, transitions, cancel, and status
New — 144 linked fieldwork jobs across primary MaxDiff, fresh repeat, five diagnostics, open ends, and preference cross-check
New — Code-only MaxDiff, stability, diagnostic, preference, and red-flag reducers with persisted source lineage
Improved — Worker output fails closed on zero respondents, missing study IDs, missing aggregates, or missing respondent-row lineage

The platform gets a home: settings, projects, skills, assets — 2026-07-21

A dedicated Settings page with full billing control, projects to organize your work, Hermes-style skills and memory, an assets gallery, grouped navigation with plain-language explanations of every tool, past runs on every tool page, live streaming replies with visible reasoning, and a stack of fixes.

New — Settings: plan, usage meter, and the Stripe billing portal — change payment method, download every invoice, or cancel, all in one place
New — Projects: folders for your work — group conversations and insights by initiative; each project carries its own context on top of account context
New — Skills: saved playbooks Mave repeats on demand, with starter templates — write once, run anytime
New — Assets: everything generated for you in one gallery — logos, research documents, insights, files
New — Every tool page now explains itself in plain language — what it is, why you'd use it, similar tools, and your past runs of it
Improved — Navigation: advanced tools grouped into Research / Audience / Data & Files / Account & Automation
Improved — Mave streams her replies word by word, with her reasoning visible while she thinks
Improved — Sign-in is your Mavera account — one login across the whole platform
Fixed — Review queue items are clickable — see exactly what's proposed (current vs new, with evidence) before approving
Fixed — New conversations appear in the sidebar instantly, with a ⋯ menu: open, quick look inside, share

The workspace becomes a real team surface: invite people with roles, share any conversation with a link (public or workspace-only), copy messages as clean text, and see exactly where every credit goes — with spend limits you set per workspace, person, project, or conversation, daily, weekly, or monthly.

New — Team page: invite by email with a role (Owner / Admin / Member / Viewer), change roles, revoke invitations — role rules enforced server-side
New — Share conversations: a link anyone can open (no sign-in) or workspace-only; revocable; read-only transcript viewer
New — Copy any message as clean text — markdown is converted, never pasted raw
New — Usage dashboard: daily spend chart, by-operation and by-person breakdowns, plan meter — every number from the metered ledger
New — Spend limits: caps per workspace, person, project, or conversation — daily, weekly, or monthly. Hitting a cap blocks new runs; chat keeps working
New — Usage meter in the composer: live plan usage and today's spend, one click from the full dashboard

Research programs: Mave works for hours — 2026-07-20

A new kind of run: give Mave a goal and a budget and she plans one step at a time — web research with cited sources, real SEO data, synthetic studies — reads what came back, and decides the next step. Hours-long programs survive page closes and finish in the background, accumulating a grounded research document as they go. The GTM Blueprint covers 12 sections from market sizing to cost-to-sell, plus a Logo Lab that generates brand directions in parallel and learns from the ones you like.

New — Programs (sidebar → Advanced): long-form autonomous research with a live accumulating document, findings timeline, and pause/resume
New — GTM Blueprint: market analysis, competitors, TAM/SAM/SOM (web-cited), traffic-share charts, startup index, buyer-style studies, paid media, SEO/AEO, viability testing, cost-to-market, support load
New — Logo Lab: parallel logo generation (up to 100 per batch), ❤️ picks steer the next batch, brand colors/fonts/notes
New — Open web research: findings cite their source URLs and carry grounding labels (real data vs synthetic study vs model estimate)
New — Attach files in Mave chat: paperclip modal + drag-drop, images/docs/spreadsheets read before answering, voice memos transcribed automatically
API — POST /api/v1/programs (stream or background), GET /api/v1/programs/[id] (findings + report), POST /api/v1/programs/[id]/logos

The sidebar reorganized around what you're trying to do, not what the system can do. Four items + an Advanced hatch — the whole engine stays one ⌘K away.

New — My Research — one unified view of everything you've made: studies, insights, audiences, conversations, audits. Searchable, filterable.
New — Clickable example prompts on the Mave landing — one click starts a real run (test messaging, find your ICP, price a product, understand an audience).
Improved — Sidebar: Mave · My Research · Workflows · Review, plus a collapsible Advanced section holding the power surfaces (Runner, Studies, Ingest, Analyze, Files, Tool Library, API keys, Docs).
Improved — Audience creation leads with Auto — describe what you have and it picks the pathway. The nine specialized pathways collapse behind one toggle.
Improved — Studies, Runner, Ingest, Analyze, Files, and Account now render inside the workspace shell — one app, one theme, no more two-layout split.
Fixed — Review badge shows the real approval-queue count (was a hardcoded placeholder).
Fixed — Standalone pages no longer flash light mode; approval-gate labels are human-readable; the Files page says Mave (not the model's name).

Company audit + eight platform features — 2026-07-07

The comprehensive one-shot company audit, and a wave of capabilities that make research continuous instead of one-off.

New — Company audit — real SEO/traffic/competitor data (Semrush) + a market frame + three built audience segments + per-segment awareness/fit fieldwork (n=50 each), fused into one scored report. Resumable steps; partial audits stay readable.
New — Autopilot — the overnight researcher. Questions Mave couldn't answer with data queue up; autopilot runs them overnight inside a credit budget you set and delivers a morning digest.
New — Trackers — re-field the same study on a cadence, trend the code-computed metrics, get alerted in chat when a number moves past your threshold.
New — Living client reports — share a study as a branded public microsite with bounded 'ask this study' Q&A. Answers come only from the study's data; revocable; question-capped.
New — Observer room — watch a focus group live, turn by turn, and pass notes to the moderator mid-session.
New — Persistent personas — named synthetic people who remember previous interviews across sessions.
New — Hallway test — a 30-second n=12 directional gut check for a handful of credits, with one-click upgrade to a full study.
New — Research plan for a budget — give Mave a goal and a credit budget, get a sequenced study mix with the arithmetic shown (costs from the measured price table, in code).
API — MCP server at /api/mcp — drive Mavera natively from Claude Code, Cursor, Cowork, or any MCP-capable agent. 173 tools with real JSON schemas.

Honest pricing + budgets — 2026-07-06

Credits re-anchored to measured compute cost, verified against live provider wire data — and spending controls at every level.

New — Public price sheet at GET /api/v1/pricing — measured typical credits per operation, USD at every plan rate, and the guarantees (pre-flight estimates, automatic refunds, per-request usage headers).
New — Workspace budgets: set a cap, a warning threshold, and top up — in the app and the API. Caps block at reservation time; you can never be charged past one.
New — Per-conversation soft/hard limits with workspace presets. Soft warns in chat; hard blocks new work while chat stays available to explain why.
New — Live budget bar at the top of every conversation, refreshed as runs settle. Mave proactively warns as thresholds are crossed — once per crossing, never nagging.
Improved — Every metered operation now bills prompt-cache writes correctly and pre-flight estimates are calibrated to measured production runs.
Fixed — Runs started from the app UI now settle credits exactly like API runs (app spend previously went unmetered).
Fixed — Pre-flight estimates apply schema defaults — a default-shaped focus group no longer estimates 0 and then charges.

Account context & memory — 2026-07-04

Mave knows your business — editable, queryable, and self-updating.

New — Context page: everything Mave knows about your business, product, customers, competitors, style, preferences. Add, edit, search, pin, delete — and 'What Mave sees' shows the exact text the model gets.
New — Auto-learning: Mave extracts durable facts from conversations, tags them LEARNED, and shows a 'Noted for next time' chip so nothing enters memory silently.
API — Full context CRUD API: GET/POST/PATCH/DELETE /api/v1/context + /api/v1/context/render (the assembled block, exactly as injected).
Fixed — Context is background knowledge, never a gate — your current request always wins over stored preferences.

Chat-first Mave + the docs site — 2026-07-02

Mave became a conversational orchestrator, and the docs became a real site.

New — Chat-first triage: 'hi how are you' gets a chat reply in seconds; research questions get a plan. The planner is a tool Mave invokes, not the front door.
New — Inline clarifications — Mave asks follow-ups as chat bubbles with tappable options. No more modals.
New — Results-aware conversation: finished studies pin to the thread as cards; every follow-up answer carries a confidence score and low-confidence answers offer the study that would actually answer it.
New — Self-hosted docs site (mavera-docs.vercel.app): sidebar nav, ⌘K search, grounded 'Ask AI' assistant, rich code blocks.
Improved — Durable cross-lambda run control — approvals and answers reach runs reliably; multi-minute studies no longer die at the function wall mid-stream.
Fixed — Honest zero-data insights: a study that returns nothing says so, instead of dressing it up.
API — GET /api/v1/quality — the public realism scorecard from the live feedback ledger. GET /api/v1/quality/chat — every low-confidence answer with the study that would improve it.

The realism loop — 2026-07-01

A built-in judge grades transcripts, gaps become repair tickets, and measured realism climbed 55 → 83.5 across the loop.

New — Judge feedback loop: a second frontier model grades every transcript; recurring gaps accumulate into a repair queue and become fixes.
New — Model-realized panels — one orchestrator call invents all panelists' prose together (+7.5 judge points over template panels, measured on live A/B seeds).
New — Code-computed diversity telemetry on every focus group: lexical overlap, length variation, opener uniqueness, residual idioms.
Improved — Fable-first model policy across all frontier tiers; persona prompts rebuilt (v2) with A/B evidence.
Improved — Study ingestion: multi-format extraction → structured dossier → analysis suite → audience-targeted fielding corroboration → self-verifying PDF report.
Fixed — Dozens of realism tells killed with code, not vibes: stimulus parroting, convergence theater, borrowed idioms, age-incoherent personas, cloned speech patterns.

Production API + conversation engine — 2026-06-29

The public REST surface hardened to production grade, and the ten conversation modalities went live.

API — Auth, rate limiting, and request logging on every v1 route; workspace scoping; API-key management (UI + CLI); interactive reference docs.
API — Billing + usage metering + members + invitations. Model spend metered on every path.
New — Ten conversation modalities live-verified: focus groups, deep interviews, debates, projective sessions, laddering, card sorts, diary studies, co-creation, media tests, website tests.
New — Insight Miner: steering, divergence detection, auto-branch — plus the find → branch → verify loop for testing a hypothesized flip-condition live.
New — Personas carry real conversational memory and current world-awareness via live web search.
New — deep_research — Mave wields live web search and sandboxed code execution inside plans.

Foundation — 2026-06-19

The platform assembled: engine clusters, governance, the designed app shell, and the registry-grounded UI.

New — 63 PRD units: audience engine (F-cluster), connectors (D), external data + indices (E), flows + templates (C), governance/provenance/approvals (H), app surfaces (G).
New — The G16 review gate: canonical changes require human approval, non-bypassable.
New — Designed app shell wired to the real registry — every surface renders real tool output, no mock data.
API — Public v1 REST + OpenAPI surface over the tool registry.