Shipping in the open
Every email about fundraising, every Drive doc about pricing, every Slack message between Sankritya, Kinjal, and Mohan flows into our own company brain. We’re validating with design partners over the next month - but right now, the live answers below come from our actual data, updating every five hours.
21 ships · last update June 5, 2026
Your deployed SKILL.md can quietly stop matching how your team actually works, and agents act on the stale version with full confidence. Lore now regenerates each skill from your current real work and, when the deployed process no longer matches, files a cited report: what the doc says, what your recent Slack threads, emails, and commits show instead, and a one-click patch. Every drift claim links to the exact source it came from.
New MCP tool: lore_skills_search. Instead of dumping the whole skill catalog into an agent's context, the agent describes the task in plain language ("how do we pick a compliance vendor") and gets back just the few most relevant skills, ranked. It loads only the winner. The brain can hold thousands of skills and the agent still pulls a tiny, exact context. Fresh skills get embedded the moment they're emitted, so they're searchable instantly.
We dogfood the product to run our company. When a new waitlist signup lands, an agent reads our onboarding runbook from Lore's own brain, researches the company, and drafts personalized outreach that cites the runbook. Our marketing posts work the same way. We review and approve; the agent never sends on its own. Same MCP a customer connects to.
Fourth source live. The brain now reads our Notion workspace alongside Drive, Gmail, and Slack. Policy pages, meeting notes, and writeups all flow in and get treated the way each kind deserves.
When the agent runs into something it doesn't know yet, the dashboard now surfaces the gap with the agent's best-guess answer. One click to approve, and next time anyone asks, the answer is there, cited.
A refund policy doesn't get useful by being chopped into three context-less sentences. Policy docs, procedures, and memos now land in the brain whole, full text preserved, retrievable as one unit. Meeting notes still get broken down where that's the right move.
Third source live. Channels you pick get read with their thread structure intact, so a question and its answer stay together instead of landing as two unrelated facts an hour apart.
Cmd+K opens a chat that picks tools (connector status, last ingest, fact growth, search), runs them in front of you, and streams a markdown answer with cited facts. Works across the whole tenant brain.
A pg_cron job fires every 5 hours, generates plausible Drive docs and Gmail threads about applying to YC, talking to design partners, and shipping product, and routes them through the same engine pipeline real connectors use. The brain visibly grows between visits.
The overview now opens with a serif italic lede - "Today, your lore learned 4 facts from Gmail and Drive" - instead of generic KPIs. Cold-start, idle-day, and active-day each get distinct copy.
Replaced the toy SVG sparklines with real charts: a 12-month calendar heatmap, a stacked-area stream graph by provider, a provider→skill sankey, and a force-directed entity graph on /people. Click a calendar day to drill into the timeline.
Skill bibliographies used to be a wall of opaque [obs:xxx] pills. Now each entry collapses into a one-line preview and expands on click - provider, date, snippet. Source chips next to facts now show the brand logo plus the provider name.
When a fact says "call this Thursday", the brief now resolves it against the email's date and TODAY, drops the event from Next if it already passed, and never reuses relative phrases verbatim.
Press ⌘K from any dashboard page. A pill at the bottom of the screen surfaces it. Submitting routes you straight to the chat with the question pre-filled and auto-sent.
Each tenant gets a dedicated MCP endpoint that Claude Code, Cursor, ChatGPT Desktop, and Codex can connect to with a long-lived `lore_…` token. Skills + facts stream into the agent from the brain.
Settings → API Keys lets admins mint a `lore_…` bearer (plaintext shown once, sha256-stored), see prefixes + last-used timestamps, and revoke individual keys. The /start/done page now has a one-click test button that probes the live MCP server with the new key.
/start/ingest now shows what the engine is reading right now - file titles landing, fact counts, threshold-aware status ("3 more facts until your first skill compiles"). Auto-advances at 1 skill or 5 facts.
A pg_cron job every 5 minutes scans every workspace, finds skills missing from the file system or with ≥3 new facts since their last emit, and recompiles the highest-priority one. No more manual emit calls.
Vercel Cron + pg_cron pull every minute, check each tenant's connectors for new items, fetch them, and hand each to the engine. The first version of the live data path: every Drive doc / Gmail thread you write shows up in the brain within ~60 seconds.
The second provider through the same machinery: OAuth, watch-list of labels/threads, fact extraction with provider chips. Identical reliability to Drive - just a different shape of input.
Five-step flow with a serif italic lede on each page: workspace → connect → pick → ingest → plug-in. Each step is one decision, not a checkbox grid.
If your team has a few years of Slack / Drive / GitHub history and you want your AI agents to inherit it, we’d love to put your company brain together with us.