BRAIN MCP
Give the keys to your company knowledge. Keep the receipts.
One endpoint, governed tools. Your agent inherits the permissions of whoever issued its key — and every access it makes is written to a record you can check.
What makes Brain the MCP server you can hand an agent?
Most MCP servers hand an agent the keys to the filing cabinet. Almost none of them write down what it opened.
Read and write, governed tools
12Search documents, read knowledge, and save useful work through tools scoped to your access. Available tools depend on your workspace and permissions.
Read · 6
- search_knowledge_graphSearch the knowledge graph for relevant facts.search_knowledge_graph
- ask_brainAsk a question using accessible knowledge.ask_brain
- list_documentsList documents available to the connected identity.list_documents
- read_documentRead an accessible document.read_document
- search_documentsFind documents by subject.search_documents
- list_filesList accessible uploaded files.list_files
Write · 3
- capture_documentSave knowledge as a governed topic document.capture_document
- update_documentUpdate a document within your permissions.update_document
- upload_fileUpload a file to the workspace.upload_file
Context · 3
- get_agent_contextRead the connected identity and its workspace context.get_agent_context
- list_skillsDiscover available workspace skills.list_skills
- list_pluginsDiscover available workspace plugins.list_plugins
Your agent inherits your access
A key acts strictly as the member who issued it, and the check runs on our side, not in the client.
Every read is recorded
Appended to a hash-chained record, designed so that tampering is detectable — and content-blind by design.
Revoke a key in one click
Rotate or revoke at any time. The next call stops.
The next call it makes is refused. Everything it already read stays in the record.
Masked before it leaves
Sensitive content is masked before it leaves our systems, so before the agent ever sees it.
EU orders may be refunded within 30 days of delivery. Escalations route to for approval, and card refunds settle to .
Masked by your rules before it leftBring your own model key
Processing runs under your own provider agreement. Nothing you connect trains a model.
Calls run under your own agreement with the provider. Nothing you connect is used to train a model.
Why not just connect each source with its own MCP?
For two sources, do. Past that, every connector is another round trip and another pile of tool definitions competing for the same context budget — so the agent spends its window deciding where to look.
The part a latency benchmark misses is governance. Eight connectors is eight permission models and eight record formats, and no single place that can answer what an agent actually read. One endpoint gives you one answer.
- 8 permission models
- 8 record formats
- 8 round trips per question
- No single answer to “what did it read?”
- 1 permission check
- 1 record, one format
- 1 call, sources resolved inside
- One place that answers it
Is this a replacement for RAG?
No, and they are not competing. RAG retrieves from an index you build and refresh, and it is hard to beat when the corpus is stable and everyone asking may see the same things.
Brain MCP gives live, permission-scoped access to the source and records what was reached. Use it when the answer depends on who is asking, or when someone will need the record later. Plenty of teams run both.
- The corpus is stable
- Everyone asking may see the same things
- You want an index you control
- The answer depends on who is asking
- The source changes faster than you re-index
- Someone will need the record later
One request, end to end
Every tool call runs the same six beats. No tool re-implements identity, permissions or recording — which is what makes the record worth anything.
It sends the question. It cannot name whose behalf it acts on.
Derived from the credential on the request, never from the arguments.
Server-side. Sources this person cannot open are filtered out before retrieval.
What survives is scanned on the way out, and masked before it leaves.
With an honest count of what was withheld, rather than a silent gap.
One entry, appended and hash-chained to the one before it.
Questions developers ask before wiring an agent in
A Model Context Protocol server that gives an AI agent access to your company knowledge through governed tools. It can search, read and save knowledge — but only by calling one of those tools, and only ever as the person whose key it carries. Routing everything through one path is what lets identity, permissions, masking and recording happen in one place across separate connections.
Sooner or later someone will ask what your agent read.
“It had permission” is not an answer. Give it the keys, and keep the receipts.
/ Get started today
