GAIP's data, in bulk.

Datasets

The weekly totals cover the Observatory (targets, reachability, response times, protocol versions), the capability record and its 2026-10 reference panel, revisit tiers and coverage, standards adoption (OAuth metadata, agent cards, MCP server cards, AI catalogues, x402, llms.txt), what agents changed that week by kind (declaration drift, protocol-version-only changes counted apart) and declared totals (tools per server, behaviour hints, output schemas, sign-in kinds, x402, registry churn), the agentic-commerce shop panel and the long-run datasets (agent keys, AI permissions, client behaviour by client family, agent terms, MCP server package history from npm and PyPI with the OSV advisories for those packages, and agent prices and ERC-8004 cards once switched on). Every column is described in the data dictionary (JSON).

Weekly downloads

Built once, when the week's snapshot is frozen, and stored; the same address always returns the same week. The newest week is always at /data/weekly/latest/<file> (a redirect). Counts from 1 to 4 are withheld and shares are given only when both counts are at least 5, as in the snapshots. No host is named in the totals files. Each file's SHA-256 is recorded with its snapshot id in the release registry. Parquet is not offered.

The first weekly snapshot covers 2026-W40 (28 Sep - 4 Oct 2026) and is frozen after that week ends, on 5 Oct 2026; its files appear here when the Observatory builds them.

Public agents GAIP watches

The public MCP servers and A2A agents GAIP's Agent Observatory watches, with their names and identifiers exactly as the public registries list them: facts as listed by those registries, not an assessment, result, score or ranking of any party, and no observation of any of them. Fields, and nothing else: GAIP's target id, the source registry, the name and identifier exactly as the registry lists them, the kind (MCP server or A2A agent card), cohort memberships and the date first watched. No results, states, scores, tools or response times. Left out: agents on GAIP's exclusion list or on a host paused by an opt-out, agents retired from their registry, agents that did not come from a public registry (hosts callers asked about, cards a caller published or captured) and GAIP's own agents.

watched-agents.csv · watched-agents.jsonl · watched-agents.json. Rebuilt daily (UTC) at the Observatory tick and after any opt-out; never computed on request.

Built 2026-10-03T07:53:20Z: 11634 agents listed. Panel AGENTS-COHORT-2026-10: 4105 members listed, 52 withheld; the registered membership hash cannot be recomputed from the list alone while members are withheld (see verification in the metadata file).

Licence: Not covered by CC BY 4.0 (which covers GAIP's aggregate figures only): the names and identifiers are as listed by public registries; GAIP's terms of use apply (https://www.gaipagents.com/terms). Opt out: an operator can opt out through corrections and opt-out; an opted-out or excluded agent is left out from the list's next build, at the latest the next day. A registry name can contain a person's name or handle (for example a GitHub username in an MCP server name): see the privacy notice.

Check the panel fingerprint and repeat the sample

The 2026-10 reference panel's registered membership hash (release registry, id AGENTS-COHORT-2026-10) is SHA-256 over the panel's target ids, sorted and joined by newlines. Take the rows whose cohorts include AGENTS-COHORT-2026-10 and hash their ids: when no member is withheld the result equals the registered hash. Members that opted out, were retired from their registry or did not come from a public registry are left out of the list, so the metadata file gives how many are withheld and why, and the hash of the listed members. A quarter's random sample uses the seed SHA-256 of the quarter label (e.g. 2026-Q4); each target id is ranked by SHA-256(seed + ":" + target_id) and the lowest ones (200) are taken.

import csv, hashlib, json
with open("watched-agents.csv", encoding="utf-8", newline="") as handle:
    rows = list(csv.DictReader(handle))
def membership_sha256(ids):           # sorted target ids joined by newlines
    return hashlib.sha256("\n".join(sorted(ids)).encode("utf-8")).hexdigest()
panel = [r["target_id"] for r in rows if "AGENTS-COHORT-2026-10" in json.loads(r["cohorts"])]
print(len(panel), membership_sha256(panel))   # compare with /v1/free/data/releases (and withheld_members)
quarter = "2026-Q4"; seed = hashlib.sha256(quarter.encode("utf-8")).hexdigest()
rank = lambda i: hashlib.sha256(f"{seed}:{i}".encode("utf-8")).hexdigest()
sample = sorted(r["target_id"] for r in rows if "AGENTS-SAMPLE-2026Q4" in json.loads(r["cohorts"]))
print(membership_sha256(sample), max(map(rank, sample)) if sample else None)

Licence

Aggregate figures (the totals files, the snapshots, the summary, the index and the monthly report) are licensed under CC BY 4.0: reuse them for any purpose, including commercially, crediting the GAIP Agent Observatory with the snapshot id. The licence does not cover per-agent or per-shop results, such as the change-log files, receipts, GAIP's software, its name or logo; GAIP's terms of use (draft) apply to those.

How to cite

GAIP Agent Observatory. GAIP Agent Observatory: weekly agent-web snapshots, snapshot AGENT-WEB-YYYY-Www. https://www.gaipagents.com/data (CC BY 4.0 for aggregate figures; facts observed at the stated time).

@misc{gaip_agent_web_agent_web_yyyy_www,
  author = {{GAIP Agent Observatory}},
  title = {GAIP Agent Observatory: weekly agent-web snapshots, AGENT-WEB-YYYY-Www},
  year = {2026},
  howpublished = {\url{https://www.gaipagents.com/data}},
  note = {Aggregate figures licensed CC BY 4.0. Snapshot id and file SHA-256 in https://www.gaipagents.com/v1/free/data/releases}
}

CITATION.cff · Croissant metadata · permanent ids

Methods

How GAIP observes, what it keeps and how anyone can check it: methodology, Agent Observatory, dataset method versions and the daily signed anchors. To correct a record or opt out: corrections and opt-out; an opted-out agent's change entries are removed from the stored change files.