Agent Layer Protocol
ALP root router
ALP is navigation in the space of meanings. An agent says what it is looking for, and the router answers with the sites where that meaning lives, ranked, with the reasons behind every score.
Early preview. The index holds a test catalogue, the protocol is a working draft, and answers may change without notice.
For agents
The router speaks the Model Context Protocol. Point an MCP client at
https://alp-geo.net/mcp
and the agent gets six tools:
alp_list_axes | the vocabulary of the index: axes, values, sites behind each |
alp_list_sites | the directory of the index: every site, host by host |
alp_query | sites for a need, ranked, each with a trust block |
alp_explain | why a site stands where it stands, or why it is absent |
alp_get_site | what a site declares and what the router found out about it |
alp_get_claim | who asserts, supports, endorses and disputes a claim |
For function calling without MCP, take the descriptor of the same tools in the format of your model and post the arguments the model returns:
GET https://alp-geo.net/v1/tools?format=openai also: anthropic, gemini
POST https://alp-geo.net/v1/tools/alp_query
Or give the agent the ALP skill: a folder with instructions and a small script. It costs the agent almost nothing until it searches, maps the user's words to the vocabulary outside the model, and asks the user before it ever sends anything back to ALP.
API
POST /v1/query | rank a query; ?explain=1 adds the breakdown of each score |
POST /v1/explain | the place of one site for a query |
GET /v1/site/<alp_id> | the indexed meta of a site and its trust tier |
GET /v1/sites | every site of the index, a page at a time; also as a page |
POST /v1/generate | a meta for a host, written from its own pages and markup; also as a page |
GET /v1/claim/<claim_id> | who asserts, supports, endorses and disputes a claim |
GET /v1/axes | the vocabulary of the index |
GET /v1/demand | what agents asked for that the index lacks; also as a page |
POST /v1/report-sources | sites an agent found after an empty answer, sent only with its user's consent |
GET /v1/health | state of the router and the size of the index |
curl -s -X POST 'https://alp-geo.net/v1/query?explain=1' \
-H 'content-type: application/json' \
-d '{
"alp_version": "0.0.0.01",
"intent": "warm winter outerwear in Montana from a verified seller",
"axes": [
{"axis": "object", "value": "retail/clothing/outerwear", "weight": 1.0, "mode": "filter"},
{"axis": "geo", "value": "us/mt", "weight": 0.95, "mode": "filter"}
],
"claim_verification": "opt-in"
}'
Queries are limited to 60 a minute per client. The router does not log the intent or
the address of the client. When a query asks for a value of the vocabulary that no site
declares, the router counts that value as demand, with the day and an
alias of the client that changes every day, so that the index grows where agents look.
"response_options": {"record_demand": false} turns that off for a query.
For site owners
A site describes itself to routers in /.well-known/alp.json. The router
reads that file with its crawler and checks the pages the file points to.
See how the crawler behaves and how to keep it away.
No file yet? Let the router write one from your pages and your schema.org markup, read it, and publish it on your host. A meta a site publishes itself scores higher than one the router wrote on its behalf.