How Brain works.
From an empty workspace to Skills your agents can execute — the whole journey in seven steps. The engine underneath is one loop: INGEST reads where work happens, DISTILL finds what recurs, SERVE hands it to any agent over MCP.
Name your workspace.
Create an account at /sign-up and name your workspace — your first lobe forms next.
A lobeis where knowledge lives. Each lobe is its own isolated scope — sources, knowledge graph, Skills, MCP endpoint, budget. Lobes in one workspace never see each other's data, so one company can keep engineering, support and a client engagement as separate brains under one login.
Describe your team.
Two sentences about what your team does — Brain generates the lobe's industry profile from them.
The profile is the lens every later stage reads through. It fixes six dimensions before any of your data is touched:
- the entity types worth tracking (people, systems, accounts, tickets…)
- the fact categories that matter in your domain
- the procedure archetypes your workflows tend to follow
- the classification labels for sorting raw events
- the sensitivity rules — what gets redacted at ingest, before storage
- a glossary of your team's own vocabulary
Generation takes a couple of minutes. You pick an intake for the lobe while it forms — which brings us to the doors.
Feed the lobe — four doors in.
Every door lands in the same pipeline: redaction at ingest, dedupe, provenance, the trust ladder.
Connect your tools
Point Brain at where work happens — Slack, GitHub, Jira, Gmail and more. Read-only OAuth in minutes: scheduled pulls, webhooks, full backfill. You choose which channels, repos or projects, and how far back. Your history stays yours.
Upload directly
No OAuth, no IT approval: paste text, or drop markdown, a zip of notes (an Obsidian vault is exactly this), a Slack export, .eml/.mbox mail, PDF, DOCX or CSV. A Slack export lands with the same identity as the live connector — upload the export for deep history, connect OAuth for freshness, and the overlap dedupes to zero. Caps: 25MB per file, 5,000 events per upload.
Watch a public community
Paste a public Discourse forum URL — zero credentials. Brain reads the public JSON API and distills the community's recurring answers the same way it treats your own threads.
Explore first
Skip intake and look around the console with the lobe empty. Every door stays open — connect or upload whenever you're ready.
The upload endpoint doubles as an ingest API, so any system without a connector can push its own exports:
curl -X POST "https://brayn.tech/api/uploads?filename=runbook.md" \ -H "Authorization: Bearer <session token>" \ -H "Content-Type: text/markdown" \ --data-binary @runbook.md
One honesty note, stated in the console too: a single-source dump plateaus around medium confidence by design — proven needs a second source or live recurrence. Upload the history, then connect a tool or capture outcomes to prove skills.
Nine stages find what recurs.
Nine stages find the procedures that recur, resolve conflicting versions and drop the noise. Five extractors work every cluster in parallel — entities, policies, procedures, decisions, cases.
This runs in the background after every sync or upload — first a full pass over the history you brought, then incrementally as new events arrive. Big drops distill in batches, so a large upload keeps producing skills over several runs. You can watch it live on the console's Pipeline page.
The trust ladder.
Nothing is fabricated, and nothing is trusted by default. Every skill earns its standing.
Every emitted Skill carries provenance— the source events it was distilled from, down to the message. In the console's Review queue each skill sits on a one-way ladder: unproven when first distilled, proving while it accumulates evidence, proven once it recurs across sources. A wrong or dangerous skill gets quarantined — pulled from serving immediately; restoring it starts back at unproven, not proven.
Confidence surfaces everywhere the knowledge does: answers in Ask cite their skills with low / medium / high bands, so a reader always knows how much weight the evidence carries.
Any agent, any model, no lock-in.
Every procedure ships as a portable SKILL.md from your lobe's MCP server — list_skills, search_brain, who_owns.
For coding agents:Console → Serve shows your lobe's MCP endpoint and mints a token. One command attaches Claude Code, Cursor, or an in-house agent:
claude mcp add --transport http brain <your MCP endpoint> \ --header "Authorization: Bearer <token>"
For everyone else:connect Brain to the chat app your team already uses. In claude.ai: Settings → Connectors → Add custom connector → paste the MCP endpoint → approve on Brain's consent page. In ChatGPT: Settings → Connectors → add the MCP server by URL → approve. From then on, ordinary chats pull your workspace's provenance-backed Skills. Approval grants read access to servable Skills and knowledge — never your source credentials, and it cannot execute anything.
No agent at all?Ask, in the console: type a question, get an answer grounded in the lobe's Skills with the sources and their trust bands shown. Same knowledge, no setup.
Grow the brain.
The console Overview draws your brain as it actually is — each lobe a region, forming as you create it.
- More lobes — one per team, per client, per community. Isolation is the point: separate sources, separate Skills, separate MCP endpoints. The free plan includes three.
- Take it with you — export any lobe as an Obsidian vault: every skill, entity and case as linked markdown notes. Your distilled knowledge is yours, in the openest format there is.
- Stay in control — Settings carries the redaction policy, retention, usage metering and a hard budget breaker per lobe. Disconnecting a source or deleting a lobe erases its data, not just its link.