Skip to main content

ARC developer documentation

Understand agent access

The model behind the endpoints: what decides whether a call succeeds, what a blocklet chooses to expose, and which surface to use for which client.

Connect a client and Authorize a client are enough to get an agent talking to a blocklet. This section explains why the results look the way they do, so you can predict them instead of probing.

Three questions organise it.

What decides whether a call succeeds

Four independent gates: the method allowlist, the credential, the path policy, and the provider's capabilities. Most surprises come from assuming a credential clears all four. It clears one.

Access tiers

What a blocklet chooses to expose

An ARC blocklet does not expose its data because it is running; it exposes what its manifest declares. The same declaration drives the MCP content tools, the llms shards, the sitemap, and the web pages. Two blocklets on the same runtime can answer very differently.

Tools for what is callable · Declare what agents see for what drives it

Which surface to use

There are three ways in, meant for clients of different capability. They serve the same data.

Client canUsePage
Speak MCP/mcpDiscovery surfaces for how to find it
Issue HTTP, but has no tool-calling harness/api/afs/rpcAFS over HTTP
Only fetch text/llms.txt and its shardsllms.txt

The text surface names the other two, so an agent that starts there can move up to a better one.

Reading order

If you are new to this board, read Access tiers first, because every other page in this section refers to its four gates. The rest can be read in any order.