Ask, and get sources back
Run the same recall your agents use, and get the material it drew on alongside the answer — so whatever you build can show its working too.
Developers
Build on a memory layer that already knows who may see what
A versioned HTTP interface with a published schema. Your software reads and writes alongside your agents — through exactly the same permission checks, with keys that cannot reach past their workspace.
55
documented endpoints
Counted from the published schema, which you can browse before writing anything.
What you can actually do with it
Not a read-only feed with a roadmap attached. Every item below is a live endpoint in the published schema today.
Drive the approval queue
List what is waiting, approve it, send it back, or act on a batch at once. If your team already has a review process, it can stay where it is.
Webhooks, both directions
Have MeshTale call you when something happens, and let your systems push content in. Signing secrets rotate, deliveries are logged, and a failed one can be replayed.
Read the audit trail
Pull the recall log and security events for a workspace, so your own dashboards can answer what an assistant was allowed to read.
Take your content out
Export a whole workspace, or a citation bundle for one answer with the sources behind it. Leaving should be as available as arriving.
Keys that cannot wander
An API key is bound to a single workspace. Not scoped by convention, or by a filter someone remembered to apply — it cannot address anything outside it.
The part that matters
Your code goes through the same checks as everything else
The usual arrangement is that an API key is an administrator — it sees everything, and keeping one customer’s data out of another’s response becomes your job, forever, in every endpoint you write.
Here the boundary sits underneath the query. A key belongs to one workspace, and there is no parameter that widens it.
$ curl -X POST /v1/search \
-H "Authorization: Bearer mk_live_…" \
-d '{"q": "renewal terms"}'
↳ results scoped to the key’s workspace
every hit carries its source
the recall is logged
There is no workspace_idoverride. Another client’s content is not filtered out of this response — it was never in scope for this key.
Three things already built on it
Two of them are ours. If the interface were not enough to build a real tool on, we would have found out first.
agent-sync
Command-line tool that pushes coding-agent sessions into a workspace as you work.
Published on npm
Obsidian plugin
Sends notes from a vault straight into the memory layer, authenticated with an API key.
Open source
The agent connection
The same recall your assistants use is two tools registered over the open protocol.
Any MCP client
Read the schema before you write anything
API access is on the free tier, so you can make a real call against your own content before deciding whether any of this fits.