Bring your own client
Connect any client that speaks the Model Context Protocol, plus direct REST access for your own code.
Connections
Make your knowledge available to the AI you choose.
Connect the AI clients you already use, authorize only the Libraries each one needs, and inspect every retrieval after the fact.
Connections · Your AI knowledge gateway2 active
2
Active connections
12s
Last call
100+
Available tools
418
Calls this month
Connections2 active
Claude Desktop
OAuth · Atlas Launch subtree
Research assistant
OAuth · read only
Local client
Revoked 3 days ago
Granted scopes
knowledge:readclaims:readposts:read
ActivityLast 7 days
Claude Desktop
Searched Atlas Launch across 8 authorized sources
12 seconds ago
Research assistant
Read Launch brief and supporting citations
8 minutes ago
Claude Desktop
Created a source-linked findings artifact
Yesterday
An illustration of the Connections workspace, using fictional data.
What you can do
Authorized access, not a blank cheque.
OAuth, tokens, or API keys
OAuth with PKCE is preferred for AI clients. Manual tokens and API keys stay available for scripts and servers.
Library-level grants
Authorize a Library subtree and its future descendants, or narrow a connection down to specific companies.
Explicit scopes
Read scopes are the default. Privacy-sensitive and write scopes are never granted automatically.
Inspectable retrieval
Every successful tool call is recorded, so you can see which client read what, and when.
Revocable at any time
Edit a connection's grants or revoke it outright. A downgrade never broadens what an existing token can reach.
Tokens are stored only as SHA-256 hashes, and a connection can never see more than you authorized.
Scoped by design
Decide exactly what each client can reach.
A connection combines two independent limits: which Libraries it can see, and which capabilities it can use. Both are chosen when you authorize the client, and both can be narrowed later.
- Subtree grants include a Library's current and future descendants.
- knowledge:read is the canonical read scope, and the only one granted by default.
- memory:read and email:read must be selected explicitly every time.
- Write scopes stay hidden unless your deployment enables them.
Add connection · Authorize scopes
Libraries this client can reach
Atlas LaunchSubtreeGranted
Competitive intelSubtreeGranted
Personal archiveSubtree—
Read scopes
knowledge:readcompanies:readclaims:readdrafts:readposts:read
memory:read and email:read are privacy-sensitive. They are excluded from default grants and must be selected explicitly.
Verifiable after the answer
See what the model actually read.
Prism does not ask you to trust a black box. The Activity tab records each accepted tool call with its client, scope, and result shape, so an answer you got yesterday can still be traced back to its sources.
- Filter activity by client, scope, or time range.
- Cross-check a retrieval against Knowledge Health for the same Library.
- Read tools are annotated read-only and idempotent.
Connections · ActivityLive
Every successful tool call
Claude Desktop · knowledge:read
search_library — 8 sources, 4 chunks returned
12 seconds ago
Claude Desktop · knowledge:read
get_agent_context_pack — profile: standard
4 minutes ago
Research assistant · claims:read
get_evidence_for_claim — 3 supporting sources
8 minutes ago
Your knowledge should outlast the tool.
Start free. Build one durable Library, authorize one AI connection, and make the context you own useful everywhere.