Open data

Data dictionary

Every column of the weekly download files and every top-level field of GAIP's main public data answers. Machine-readable: /v1/free/data/dictionary. Back to open data.

totals.csv and totals.jsonl (CC BY 4.0)

One row per ISO week, dataset and metric.

Field	Type	Description
`snapshot_id`	string	Citable id of the weekly snapshot, AGENT-WEB-YYYY-Www.
`week`	string	ISO week the figures cover (Monday 00:00 UTC to the next Monday), e.g. 2026-W40.
`dataset`	string	Which record the figure comes from: a section of the weekly snapshot (observatory, capability_record, revisit_tiers, reach_coverage, market_coverage, standards_adoption, shop_panel, declaration_drift: what agents changed that week by kind with protocol-version-only changes apart, declared_totals: totals of declared tools, behaviour hints, output schemas, sign-in kinds, x402 and registry churn, declared_derivations: share of agents asking for each input-field category, agents pointing to outside domains, operator concentration; domain names are not in the files) or one of the long-run datasets (agent_keys, ai_permissions, client_behaviour, agent_terms, agent_prices, erc8004_cards, package_history: weekly npm and PyPI facts of MCP server packages and the OSV advisories for them).
`metric`	string	Path of the figure inside its dataset block, dot-separated (as in the snapshot JSON at /v1/free/ecosystem/snapshots?week=). A share cell gives two rows: <cell>.count and <cell>.share.
`value`	number	The figure. Empty when withheld (see withheld) or not available that week.
`denominator`	number	For count and share rows of a share cell: the number of items the share is taken over (the cell's 'of'). Empty otherwise or when withheld.
`unit`	string	count, share (a fraction from 0 to 1), milliseconds, seconds, hours, days or number.
`withheld`	boolean	true when GAIP withheld the value: a count from 1 to 4, or a share whose count or denominator is under 5 (minimum cell size 5).
`method_version`	string	Method version(s) of the dataset in force that week, from /v1/free/observatory/methods (several joined with +); for snapshot sections without a methods entry, the snapshot schema version.

changes.csv and changes.jsonl (not CC BY)

Not covered by CC BY 4.0: per-agent entries of GAIP's public change log, exactly as served at /v1/free/observatory/changes; GAIP's terms of use apply (https://www.gaipagents.com/terms).

Field	Type	Description
`snapshot_id`	string	Citable id of the weekly snapshot, AGENT-WEB-YYYY-Www.
`week`	string	ISO week in which GAIP observed the change.
`sequence`	integer	Position of the entry in GAIP's change log.
`target_id`	string	GAIP's id of the observed public MCP server or A2A agent (obs_...).
`target_kind`	string	MCP_REMOTE, A2A or another target kind.
`target_locator`	string	The public endpoint or agent card URL GAIP read, as the public change feed shows it.
`before_hash`	string	SHA-256 of the declared document before the change.
`after_hash`	string	SHA-256 of the declared document after the change.
`observed_at_utc`	string	When GAIP observed the change (RFC 3339, UTC).
`diff`	object	What changed, as the public feed shows it (tool names, hashes, short excerpts). JSON text in the CSV.
`labels`	array	Change classes (for example tool_added, description_changed). JSON text in the CSV.
`receipt_id`	string	GAIP receipt for the observation, when one was issued.
`receipt_url`	string	Where to verify that receipt.
`previous_entry_sha256`	string	Hash of the change-log entry before this one (the chain).
`entry_sha256`	string	SHA-256 of this entry; commits to previous_entry_sha256.
`evidence_class`	string	What kind of evidence the entry is (a fact GAIP observed, not a verdict).

watched-agents.csv and watched-agents.jsonl (not CC BY)

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). 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.

Field	Type	Description
`target_id`	string	GAIP's id of the watched agent (obs_ + 24 hex characters).
`source_registry`	string	The public registry that lists it: official-mcp-registry or a2aregistry-org.
`registry_name`	string	The name as that registry lists it (the Official MCP Registry's title; empty when the registry gives none, and for the A2A Registry, from which GAIP keeps no name).
`registry_identifier`	string	The identifier as that registry lists it: the Official MCP Registry's server name (e.g. io.github.owner/server) or, for the A2A Registry, the URL of the agent card it lists (wellKnownURI).
`kind`	string	MCP_REMOTE (a remote MCP server) or A2A_CARD (an A2A agent card).
`cohorts`	array	Cohort memberships: AGENTS-COHORT-2026-10 (the reference panel), AGENTS-COHORT-YYYYQn (first watched in that quarter) and AGENTS-SAMPLE-YYYYQn (that quarter's random sample). JSON text in the CSV.
`first_watched_date`	string	The date (UTC, YYYY-MM-DD) GAIP first watched it.

reliability-index.csv and reliability-index.json (CC BY 4.0)

One row per period (final month or the current 30 days), cohort and protocol; built from the copies the Observatory tick stores (final months frozen once; current hourly).

Field	Type	Description
`period`	string	YYYY-MM for a final month, or rolling_30_days for the current block
`status`	string	final or provisional
`window_start_utc`	datetime	Start of the period (UTC)
`window_end_utc`	datetime	End of the period (UTC, exclusive)
`cohort`	string	all (every public agent GAIP watches) or panel (AGENTS-COHORT-2026-10)
`protocol`	string	MCP, A2A or ALL
`agents`	count	Agents with at least one visit (empty when under 5: the cell is withheld)
`visits`	count	Visits: observations in which GAIP sent a request
`visits_reachable`	count	Visits where the host answered (any HTTP status); empty under 5
`visits_reachable_share`	share	visits_reachable / visits
`agents_reachable_95`	count	Agents reachable on at least 95% of their visits; empty under 5
`agents_reachable_95_share`	share	agents_reachable_95 / agents
`response_time_p50_ms`	milliseconds	Median response time of successful handshakes
`response_time_p95_ms`	milliseconds	95th percentile response time of successful handshakes
`successful_handshakes`	count	Reachable visits with an HTTP 2xx answer and a response time

Units

count: a number of items
share: a fraction from 0 to 1 (count / denominator)
milliseconds: time in ms
seconds: time in s
hours: time in h
days: time in days
number: another numeric value

Minimum cell size 5: counts from 1 to 4 are withheld.

Public data answers

`/v1/free/observatory/summary`: Agent Observatory summary (updated each Observatory cycle)

Field	Description
`schema_version`	Version of this document's shape.
`generated_at_utc`	When the summary was built (UTC).
`notice`	What the figures are and are not (facts observed at a stated time; not an assessment).
`neutrality`	No ranking, league table or score of any party; small counts suppressed.
`targets`	Public targets GAIP watches: total, by kind and by source, and those withdrawn or unpublished.
`coverage`	Targets with an observation and the share observed in the last 48 hours.
`status_distribution`	Targets by their latest observed state (counts).
`manifest_changes`	Count of published manifest changes in the last 7 days.
`ecosystem_conformance`	Aggregate counts of declared-metadata checks by protocol (cells of at least 5).
`metadata_signal_pattern_rates`	Per signal pattern: present, not present, not evaluable, and the rate.
`overall_verdicts`	Always null: GAIP issues no overall verdict.
`ranking`	Always null: GAIP does not rank.
`capability_record`	The 2026-10 reference panel (size, membership hash, status counts), change-log head and the quarter's random sample.
`tiers`	Revisit tiers: daily-tier and census sizes, cadence and coverage.
`reach`	Weekly coverage: share of known public remote agents observed in the last 7 days.
`limits`	Request budget and revisit plan (per cycle, per hour).
`cycle_health`	Scheduler state, last cycle and errors (operational).
`sources`	Registries and lists GAIP reads, with their terms basis.
`registry_archive`	Counts of archived Official MCP Registry entries (versions kept, names).
`declared_extras`	What agents declared beyond their tools (MCP instructions present, MCP Apps UI, tasks, tool-list change notices, extensions; A2A AP2, x402 and UCP extensions; 401 challenge schemes): counts over public targets, 1 to 4 shown as '<5'.
`declared_derivations`	Derived from what agents declared (data5): data_appetite (share of agents whose tools' input fields ask for each category: payment card, bank account, email, phone, postal address, precise location, government id, health, date of birth, credentials, biometric, unbounded free text, file upload; counted by keyword rules, never field names), dependencies (agents pointing to outside domains; well-known service domains shared by at least 5 agents named, others only counted) and operators (operators counted, share of agents run by the 10 largest, operators by agents run; no operator named). Cells under 5 as '<5'; writing: when the newest record was written.
`shop_panel`	The agentic-commerce shop panel's cohort and aggregate figures.
`unique_datasets`	Heads and counts of the long-run datasets (agent keys, AI permissions, client behaviour, agent terms, agent prices, ERC-8004 cards, MCP package history), and whether each own-reader dataset's collector is on.
`outcome_ledger`	Head of GAIP's monthly outcome ledger.
`weekly_snapshots`	Head of the weekly snapshot chain: hash, count and newest week.
`reliability_index`	When the reachability and response-time figures were last stored, and the newest final month's headline (share of visits reachable, p50 and p95 response time).
`retention_days`	Retention per record (null means no deletion date).
`retention_notes`	Plain-English retention rules.
`third_party_scanners`	Outside scanners named in the method and whether GAIP ran them (NOT_RUN).
`endpoints`	URLs of the per-target, lookup, change and capture routes.
`price_gbp`	Always 0: free.
`automatic_evidence_credit`	Always false.
`served`	Present when served from the copy the Observatory stored: when it was built and whether it is stale.
`license`	Licence block: CC BY 4.0 for aggregate figures, with what it covers.

`/v1/free/ecosystem/snapshots`: Weekly ecosystem snapshots (frozen, hash-chained)

Field	Description
`schema_version`	Version of this document's shape.
`title`	Name of the series.
`notice`	What the figures are and are not.
`chain`	Check of the snapshot hash chain.
`method`	Method page.
`citation`	How to cite one snapshot.
`licence`	Licence note (CC BY 4.0 for aggregate figures).
`licence_url`	Licence URL.
`corrections`	Corrections and opt-out route.
`price_gbp`	Always 0.
`automatic_evidence_credit`	Always false.
`weeks`	Index: each frozen week with its snapshot SHA-256, the previous week's and its URL.
`first_snapshot`	When the first snapshot is frozen.
`week`	With ?week=: the ISO week.
`snapshot`	With ?week=: the frozen aggregate snapshot.
`previous_sha256`	With ?week=: the previous week's snapshot hash.
`snapshot_sha256`	With ?week=: this snapshot's SHA-256 (commits to previous_sha256).
`error`	WEEK_NOT_FROZEN when the asked week has no snapshot.
`license`	Licence block (see /v1/free/data/dictionary).

`/v1/free/observatory/methods`: Dataset methods: versions and changes

Field	Description
`schema_version`	Version of this document's shape.
`title`	Name.
`meaning`	How method versions are kept and applied.
`method_doc`	Link to the method section.
`overlap_rule`	The old-and-new overlap a later method change must run.
`from_basis_note`	How each version's from_utc was set.
`datasets`	Per dataset: title, cycle, method versions with dates, what changed and known defects.
`url`	This document's URL.
`license`	Licence block (see /v1/free/data/dictionary).

`/v1/free/agent-web-index`: Agent Web Index (monthly)

Field	Description
`schema_version`	Version of this document's shape.
`name`	Index name.
`title`	Title.
`month`	Calendar month (YYYY-MM).
`status`	provisional or final.
`window_utc`	Start and end of the month (UTC).
`weeks`	ISO weeks counted in the month.
`generated_at_utc`	When the index was built.
`figures`	The headline figures, each with count, denominator and share (withheld under 5).
`declared_derivations`	Beside the figures: the month's latest frozen weekly declared_derivations block (data appetite, outside dependencies, operator concentration), as frozen; null before a week with it is frozen.
`sample_frame`	The reference panel (size, membership hash) and the registry size it is read against.
`rules`	Counting rules.
`not_collected_yet`	Figures whose collector is switched off.
`minimum_cell_size`	5: counts and totals under it are withheld.
`citation`	How to cite the month's index.
`json`	This document's URL.
`reproduce`	The published queries and the script that recompute each figure (/v1/free/agent-web-index/queries, /v1/free/agent-web-index/reproduce.py).
`weekly_totals_rule`	How each figure's count and total split into ISO-week parts, and when parts are withheld (a part under 5 withholds all parts of that figure).
`licence`	Licence note.
`licence_url`	Licence URL.
`available`	false when no copy is stored for the month.
`message`	Why not available.
`meaning`	What a missing month means.
`error`	MONTH_FORMAT for a bad ?month=.
`example`	A valid month.
`reliability`	The month's reachability and response-time figures (final once the month is frozen, else the latest 30 days): share of visits reachable, share of agents reachable on at least 95% of visits, p50 and p95 response time, by protocol and cohort.
`served`	When the stored copy was built and whether it is stale.
`license`	Licence block (see /v1/free/data/dictionary).

`/v1/free/data/releases`: Permanent dataset ids

Field	Description
`schema_version`	Version of this document's shape.
`title`	Name.
`meaning`	What an id is and that it is never edited.
`licence`	Licence note.
`licence_url`	Licence URL.
`notice`	What the figures are.
`chain`	Check of the committed registry's hash chain.
`entries`	Cohort and quarterly release ids (committed registry), each with its content and SHA-256.
`dois`	DOIs recorded for issued releases.
`method`	Method page.
`price_gbp`	Always 0.
`automatic_evidence_credit`	Always false.
`weekly_snapshots`	Weekly snapshot ids AGENT-WEB-YYYY-Www with each download file's SHA-256 (database registry, append-only, hash-chained).
`license`	Licence block (see /v1/free/data/dictionary).