Give your agent fleet a shared brain
Brainz is a product-knowledge graph built for agent orchestration. Features, decisions,
constraints, and patterns live as markdown concepts wired into a graph — and before your
orchestrator delegates a coding task, one get_context call over MCP returns a
token-budgeted markdown briefing to paste straight into the subagent's prompt. Subagents get
exactly the product knowledge they need. The orchestrator's context stays clean.
Brainz is a Brownmatter product — hosted at brainz.brownmatter.ai, or run it
locally from a folder of markdown.
A knowledge base agents can actually use
Not a wiki that goes stale in a drawer — a queryable graph whose primary consumer is an orchestrator agent, with humans as welcome guests.
Concepts, linked into a graph
Every concept is a markdown note — feature, decision, constraint, pattern, or glossary
term. Inline [[wiki-links]] and typed relations like depends_on, part_of, and constrains wire notes
into a directed graph, so context assembly follows the structure of your product instead
of guessing from keywords. And code_areas anchor each concept to the code it
describes — a briefed agent learns the rules and where to apply them.
---
title: Lead Scoring
type: feature
tags: [crm, ranking]
code_areas: [src/scoring/**]
relations:
depends_on: [contact-model]
governed_by: [rate-limits]
---
Scores leads 0-100. Recomputed on every
[[contact-model]] change; see [[scoring-decision]].Remote MCP in one snippet
Point any MCP client at your hosted vault — Claude Code, or anything that speaks
streamable HTTP. Create an API key in the dashboard, add one entry to .mcp.json, and every vault tool is live — from search_concepts to take_task.
{
"mcpServers": {
"brainz": {
"type": "http",
"url": "https://brainz.brownmatter.ai/mcp",
"headers": { "Authorization": "Bearer bz_..." }
}
}
}Hybrid search, graceful fallback
search_concepts fuses full-text search with embedding similarity via
reciprocal-rank fusion; get_context then expands through the graph to pull in
the neighbors that matter. If embeddings are unavailable — offline, no key, no model —
search degrades to keyword-only. Nothing crashes, ever.
Local-first & open source
The same brain runs from a folder of markdown on your machine: npx brainz init,
edit notes in any editor, and serve them to your agents over stdio MCP with the CLI. The
index is derived SQLite, rebuilt from your files — your knowledge stays plain markdown,
in git, forever portable.
Give your fleet a work queue
Tasks are concepts too. create_task turns a finding into first-class work —
wired to its domain, the specs that govern it, and the code areas where it lands. One take_task call claims the next open task atomically and returns it
with an auto-assembled briefing, so the agent that picks up the work starts already
briefed. You watch it happen on a real-time Kanban board: Open, Taken, Done.
Share one brain across humans and agents
Invite teammates by email — single-use links, owner / editor / viewer roles, enforced on every REST route and MCP call. Your colleagues
browse the same graph in the dashboard that your agents query over MCP. One vault, one
source of truth, everyone briefed from it.
No signup form required — your agent does it
An orchestrator can provision its own brain: hit https://brainz.brownmatter.ai/mcp with no
credentials and the endpoint exposes two public tools — signup and sign_in. One signup call creates a provisional workspace: a
first vault and an API key that works immediately, no email required. The response
contains no URL of any kind — that's the security story. The workspace expires in 7 days
unless claimed: the claim tool emails your human an approval link the agent
never sees. No passwords anywhere — Brainz is passwordless. Or skip the wiring entirely:
tell your agent to read the setup flow and it does the
rest.
# connect to https://brainz.brownmatter.ai/mcp with NO auth header → two public tools
→ tools/call signup { "vaultName": "Acme brain" }
← { "apiKey": "bz_…", "vaultSlug": "acme-brain",
"expiresAt": "…", "claimEmailSent": false }
# no URL of any kind in the response — approval links only ever go to the human's inboxThree steps to briefed subagents
Capture concepts
Point your agent at an existing codebase: the onboard tool hands it a
playbook for distilling your product into the graph. Then let agents keep capturing
knowledge as they work with upsert_concept. No account yet? Your agent can sign itself up first.
Wire the MCP
One entry in .mcp.json: hosted over HTTP with an API key, or local over
stdio. Your orchestrator sees the full toolset — search, traversal, writes, a work
queue, and the flagship, get_context.
Brief every subagent
Before delegating, the orchestrator calls get_context and injects the
returned markdown bundle into the subagent's prompt. Curated product knowledge in;
context pollution out.
Free. Actually free.
Hosted vaults at brainz.brownmatter.ai are free for everyone — no tiers, no trial
clock, no credit card, no seat limits. And because the engine is open source and every concept
is plain markdown, you're never locked in: the same vault runs from a folder on your laptop.