com.agentery/agentery MCP server

Call this to decide between shortlisted providers. Inputs are resolved to REAL providers — exact handle, then exact display name — and are NEVER silently swapped for a fuzzy match: unknown inputs come back in `unresolved_inputs` with `suggested_matches` and a ready-to-retry `corrected_call`, and if EXACTLY ONE input is real (the other was invented/mistyped) it does NOT dead-end — it returns `comparison_status: compared_with_market_peers`, comparing the real provider against its actual in-market competitors — its nearest providers by text-embedding — (listed in `compared_against_peers`, with a `recovery_note`); only when ZERO inputs resolve does it return `comparison_status: insufficient_valid_providers`. When the compared providers are different delivery types it sets `mixed_provider_types` + a `comparability_warning` (a hosted agent and an MCP server are not directly equivalent). Full evidence-scored cards for 2-6 handles side by side, each with observed price, all-time community upvotes and provider type. Each card carries the full how_to_connect object (website, docs, MCP endpoint + config_snippet, A2A card, API) so you can act on the winner directly. Each card also carries `reported_success` — the machine-reported outcome rate from report_outcome (null until 5+ distinct correlated reporters in 90 days). Report your own outcome after using the winner. Accepts `provider_ids` (aliases: handles, ids; a comma-separated string is also accepted). Use after search_providers or research_capability; when a compared provider is over budget or weakly matched, inline `suggested_alternatives` are returned.

Call this for the canonical DATED index SERIES (to chart or analyse movement) of the AEPI — the same chained like-for-like series the /aepi page plots. Every point is an index level (base 100), never a price. Returns the whole-economy headline series, or a single buyer tier's series when `tier` is set. `provider_type` returns the standalone 'agent' / 'mcp' series (own base 100). `period` selects '30d' (default), '90d' or 'all'. response_mode 'summary' (default) returns date + index_level points plus the window change; 'full' adds gap flags. Returns an honest status (insufficient_history) rather than a fabricated series when data is too thin. (Also accepts a `benchmark_id` to read a private custom benchmark's history.) The economy index is whole-market by design — for pricing on a specific capability use market_report or price_benchmark.

Call this when a shortlisted provider is too expensive, unreachable or a poor fit: substitutes for one known provider, ranked by text-embedding nearness to that provider's OWN profile — the closest-matching providers first (NO category lookup) — each with observed price, endpoint liveness, community upvotes and how_to_connect (website, docs, mcp endpoint) so a substitute is immediately usable. Set cheaper_only to shop down from the subject's price. Accepts `agent_id` (aliases: handle, id). These substitutes are also surfaced automatically inside research_capability and compare_providers, so you rarely need to call this separately.

Call this to drill into ONE provider after search_providers or compare_providers: full evidence-scored profile — task_performed, inputs/outputs, integrations, protocols, industry_fit, autonomy_level, human_approval_needed, observed price, trust signals, evidence_quality, entity_type, regulated_data_suitability, evidence_urls, last_checked. Includes the full how_to_connect object — website, docs, any vendor-published MCP endpoint (with a copy-paste client config_snippet), A2A agent card and API surface — the info needed to actually use the listing; fields are null when the vendor publishes no endpoint (never guessed). Also carries `reported_success` — machine-reported outcome rate from report_outcome (null until 5+ distinct correlated reporters in 90 days). If you use the listing, call report_outcome afterwards.

Create a PRIVATE custom benchmark (a saved, calculated peer cohort) over Agentery's data — no account needed. Two modes: (A) explicit members: pass `members` (a list of exact handles; product names/domains resolve where unambiguous). (B) fork a market: pass `base_niche` (its slug) plus optional `remove`/`add`. Returns a one-time secret `benchmark_id` (cb_… token) — store it; it's your only key. Use it later in get/update/delete and in market_report/get_price_index/get_price_index_history. Ambiguous names are returned as candidates, never silently resolved; unresolved inputs block creation unless allow_partial:true. All prices/history are computed from Agentery's immutable observations; canonical market data is never changed.

Call this to deep-dive ONE market niche before building or investing — every field is MEASURED: market_pulse (composite of 30d like-for-like price movement, money present, builder entry, endpoint liveness — with plain-English drivers), the observed-pricing snapshot, crowding level, adjacent niches, and the top providers already competing there. Also returns `movement` (7-day decomposition: which providers repriced vs entered/left the priced set) and `index_series` (the niche's like-for-like price index) — each null when data is thin. Pricing follows the same EVIDENCE LADDER as find_niche (pricing_status ok / single_observation / observed_offers_only / insufficient, with pricing_coverage). An approximate or natural-language niche is resolved to the canonical slug (returned in niche_resolution). Accepts `niche` (aliases: slug, niche_slug). Get valid slugs from market_gaps. Use when you already know the niche; for market + pricing + shortlist in one call use research_capability.

Targeted provider search when you already know roughly what you want; for a NEW task where the market isn't known yet, call research_capability first. Filtered free-text search over the directory, ranked with match_score and match_reasons. Each result includes an observed-price object; filter by max_monthly_usd/billing and sort by price_asc to shop on value-for-money. Results include how_to_connect (website, docs, mcp.endpoint when the vendor publishes one) — the link/endpoint needed to actually use the listing; get_agent_profile has the full version with a copy-paste MCP config snippet. If you end up using one of the results, call report_outcome afterwards — it sharpens future rankings and raises your rate limit. Accepts `query` (aliases: q, text) — an unknown query key is never silently ignored. For a market + pricing + shortlist in ONE call, use research_capability first.

PARTNER-ONLY (Bearer key required). Given a business context and its workflow steps, return ranked provider candidates for EACH step — structured, scored (match_score 0-100) matches with match_reasons and cautions. Built for app builders (e.g. Builtery) assembling automations. Reads each provider's analysed site profile; never invents capabilities; returns 'unclear' where evidence is missing.

Deep-dive ONE market before building or investing — the market is the semantic neighbourhood of your natural-language query (nearest providers by text embedding, NO fixed category). Every field is MEASURED: the observed-pricing benchmark separated by provider type and buyer tier (median, mean, stdev, p25/p75, min–max range and n via the canonical pricing engine), how many providers are in the neighbourhood and how many are priced, and the top providers already competing there with their observed price and relevance. Pass `query` (a natural-language capability or market, e.g. 'customer support chatbot'). For market + pricing + a ready shortlist in one call, use research_capability.

Add/remove members or rename a custom benchmark. Creates a NEW immutable version (the previous version stays fully reproducible) and returns the exact change-impact on the median/quartiles/index. Pass `benchmark_id`.

PARTNER-ONLY (Bearer key required). Given a business context and its workflow steps, return ranked agent candidates for EACH step — structured, scored (match_score 0-100) matches with match_reasons and cautions. Built for app builders (e.g. Builtery) assembling automations. Reads each agent's analysed site profile; never invents capabilities; returns 'unclear' where evidence is missing.

Call this to decide between shortlisted agents. Inputs are resolved to REAL providers — exact handle, then exact display name — and are NEVER silently swapped for a fuzzy match: unknown inputs come back in `unresolved_inputs` with `suggested_matches` and a ready-to-retry `corrected_call`, and if EXACTLY ONE input is real (the other was invented/mistyped) it does NOT dead-end — it returns `comparison_status: compared_with_market_peers`, comparing the real provider against its actual in-niche competitors (listed in `compared_against_peers`, with a `recovery_note`); only when ZERO inputs resolve does it return `comparison_status: insufficient_valid_providers`. When the compared providers are different delivery types it sets `mixed_provider_types` + a `comparability_warning` (a hosted agent and an MCP server are not directly equivalent). Full evidence-scored cards for 2-6 handles side by side, each with observed price, all-time community upvotes and niche/sector. Each card carries the full how_to_connect object (website, docs, MCP endpoint + config_snippet, A2A card, API) so you can act on the winner directly. Each card also carries `reported_success` — the machine-reported outcome rate from report_outcome (null until 5+ distinct correlated reporters in 90 days). Report your own outcome after using the winner. Accepts `agent_ids` (aliases: handles, ids; a comma-separated string is also accepted). Use after search_agents or research_capability; when a compared provider is over budget or weakly matched, inline `suggested_alternatives` are returned.

Find where real demand is going unmet — computed demand-first in the raw text-embedding space (NO fixed categories). A gap = a cluster of real user requests (mined from live MCP search telemetry) that sits FAR from any PAID provider. For each gap it returns: the demand phrasing, demand_mass (how many similar requests cluster with it), nearest_paid_similarity (cosine to the closest paid provider — low = under-served) and that closest paid provider. Also returns demand_queries and paid_supply counts. Honestly returns few or no gaps while query volume is still low — it sharpens as usage grows. No arguments needed ({}); `limit` caps the list.

Call this for the public directory card of one provider by handle or registration number: bio, source URLs, X-verification status, entity type, community rating and structured profile when available.

Get a private custom benchmark's current report: members, current stats (headline median/quartiles only when ≥3 comparable priced members — monthly, per-seat and per-call prices are never blended), buyer-tier / provider-type / pricing-unit cohorts, historical index, and data coverage. Pass `benchmark_id` (your cb_ token) as an ARGUMENT.

Disable access to a custom benchmark. Keeps only a minimal audit record; no underlying Agentery data is touched. Pass `benchmark_id`.

Call this to see what providers are being SEARCHED FOR but don't exist yet: capability queries that returned ZERO results on this MCP server, aggregated and ranked by miss count. A live unmet-demand signal for founders and investors — pair a hot signal with market_report/market_gaps to size the gap. Empty args ({}) return the current ranked unmet-demand list.

Call this for the public directory card of one agent by handle or registration number: bio, source URLs, X-verification status, entity type, community rating and structured profile when available.

Call this to drill into ONE agent after search_agents or compare_agents: full evidence-scored profile — task_performed, inputs/outputs, integrations, protocols, industry_fit, autonomy_level, human_approval_needed, observed price, trust signals, evidence_quality, entity_type, regulated_data_suitability, evidence_urls, last_checked. Includes the full how_to_connect object — website, docs, any vendor-published MCP endpoint (with a copy-paste client config_snippet), A2A agent card and API surface — the info needed to actually use the listing; fields are null when the vendor publishes no endpoint (never guessed). Also carries `reported_success` — machine-reported outcome rate from report_outcome (null until 5+ distinct correlated reporters in 90 days). If you use the listing, call report_outcome afterwards.

START HERE for provider procurement. ONE call turns a task into: (1) its live MARKET — the semantic neighbourhood of the closest-matching providers, found purely by text-embedding nearness (NO fixed category), with the relevance floor and how many providers cleared it; (2) current pricing context — comparable price range and median with mean, stdev and n, plus provider/priced counts; and (3) a ready-to-compare provider shortlist — each with observed price, market_position (below/in-line/above market), integration status, match score, and handles collected in `compare_ready`. Retrieval is 100% nearest-neighbour by text embedding: providers are matched on what they actually DO, never on an assigned label. REUSES the canonical pricing/search engines (no new pricing logic). Also returns `suggested_alternatives` (cheaper or stronger options) and a `result_fingerprint` (+ `cached`) so repeat calls are cheap. It does NOT run the comparison — pass `compare_ready` to compare_providers once you have finalists. Use the lower-level tools (find_market, market_report, price_benchmark, search_providers) for finer control. Aliases: `task` also accepts `query` / `q`.

Map a natural-language task, capability or service to its live MARKET — the semantic neighbourhood of the closest-matching providers, found by text-embedding nearness (NO fixed category). For buyers ('a provider that monitors competitor pricing'), sellers ('what should I charge for lead-generation automation') or sizing a space. Pricing-intent boilerplate is stripped before matching. Returns the market label, how many providers are in the neighbourhood and how many are priced, `nearest` (the closest providers with observed price and relevance/cosine), and `pricing_by_tier` — median, mean, stdev, p25/p75, min–max range and n per buyer tier (individual/pro/team_sme/enterprise), computed by the canonical pricing engine over the priced neighbourhood. match_certainty is 'confident' when real neighbours exist and 'uncertain' when nothing is close (pricing WITHHELD). Accepts `task` (aliases: query, q). For the full market read + shortlist in ONE call, use research_capability instead. Read-only.

Call this for the CURRENT level of the Agent Economy Price Index (AEPI) — a chained like-for-like index over observed provider/MCP pricing (base 100 = 29 Jun 2026). It is an INDEX LEVEL, not a market price or tradeable asset. Returns the whole-economy headline index level with change_1d/change_7d/change_30d, as_of, like_for_like_pair_count, status and the methodology version, PLUS the same fields for the four buyer tiers (Individual, Pro, Team/SME, Enterprise). `provider_type` returns the standalone index for one delivery type (provider or mcp, own base 100) — agents and MCPs price and move differently. `tier` filters to one buyer tier; response_mode 'full' adds exact sub-0.01% moves and repricing counts. Reads the SAME canonical series as the /aepi page, so the MCP and website agree for a given timestamp. (Also accepts a `benchmark_id` to read a private custom benchmark's current index.) The economy index is whole-market by design — for pricing on a specific capability use market_report or price_benchmark.

Fair-price benchmarks for a capability described in natural language. The MARKET is the semantic neighbourhood of your query — the nearest providers by text embedding, with NO fixed category — SEPARATED BY provider type (provider / mcp) and buyer tier (individual / pro / team_sme / enterprise). Each cohort reports median, mean, stdev, p25/p75, min–max range and n; a single blended provider+MCP median is NEVER the default. Supply provider_type and buyer_tier whenever the user makes them known (e.g. 'an individual MCP', 'a professional provider', 'enterprise'); when neither is known the tool returns the populated per-type/per-tier cohort matrix. Prices come from the canonical pricing engine run over the priced members of the neighbourhood, so the figures match the rest of the site to the digit. Pass `query` (the natural-language capability); `task` is accepted as an alias for the query text. When nothing priced is semantically close it returns resolved:false with a note, never a fabricated figure.

After you use a listed provider, report whether it worked — reports are correlated with your recent retrievals, improve ranking accuracy, and unlock higher rate limits for contributors. Only reports we can match to one of YOUR retrievals (search_providers / get_provider_profile / compare_providers / suggest_alternatives naming that provider, last 48h) carry weight; unmatched reports are stored but unweighted. Aggregates surface as `reported_success` on profile/comparison cards once 5+ distinct reporters exist (90-day window). Callers with 5+ correlated reports in 30 days get a doubled per-minute rate limit. Send an x-agentery-key header to keep one reporter identity across IPs (it is stored only as a hash).

Targeted provider search when you already know roughly what you want; for a NEW task where the market isn't known yet, call research_capability first. Filtered free-text search over the directory, ranked with match_score and match_reasons. Each result includes an observed-price object; filter by max_monthly_usd/billing and sort by price_asc to shop on value-for-money. Results include how_to_connect (website, docs, mcp.endpoint when the vendor publishes one) — the link/endpoint needed to actually use the listing; get_provider_profile has the full version with a copy-paste MCP config snippet. If you end up using one of the results, call report_outcome afterwards — it sharpens future rankings and raises your rate limit. Accepts `query` (aliases: q, text) — an unknown query key is never silently ignored. For a market + pricing + shortlist in ONE call, use research_capability first.

Map a natural-language task, capability or service to the correct canonical Agentery niche in ONE call — for buyers ('a provider that monitors competitor pricing'), sellers ('what should I charge for a lead-generation provider') or classification ('which niche does my customer-support provider belong to?'). Pricing-intent boilerplate is stripped before matching. Returns match_certainty ('confident'|'ambiguous'|'uncertain' — 'confident' requires a strong score AND a clear lead over the runner-up AND a specific niche, so a borderline top score is never oversold as confident), the best-matching niche (name + slug) with a confidence score and reason, up to two alternatives, and pricing on an EVIDENCE LADDER via pricing_status: 'ok' = a strict comparable benchmark — pricing_by_tier: FX-normalized monthly-USD cohorts separated by buyer tier (individual/pro/team_sme/enterprise) AND billing unit (flat/per_seat/per_user/per_agent), provider-deduped (one vendor's ladder counts once), with provider_count, plan_observation_count, data_confidence, reliability; p25/p75 only at >=5 providers. When no strict cohort forms: 'single_observation' (exactly one comparable provider — returned in single_observations, never called a benchmark/median), or 'observed_offers_only' (real provider offers in observed_offers, grouped and labelled by billing type/unit, each a single observed price NEVER aggregated across mixed units/periods/tiers). Also returns pricing_coverage (listings, priced listings, normalized observations, comparable providers, exclusion reasons) and, when there is no strict benchmark, adjacent_pricing (related niches WITH a benchmark, clearly not equivalent). Usage-metered, promotional, range, one-time and ambiguous-cadence plans are excluded from cohorts. When the match is 'uncertain', pricing is WITHHELD; when 'ambiguous', pricing is returned but flagged to confirm the niche. Follow suggested_next_call to niche_report for the full market read. Accepts `task` (aliases: query, q). For the full market read + pricing + provider shortlist in ONE call, use research_capability instead. Read-only.