Skip to main content
Glama
443,491 tools. Updated 2026-08-11 13:28

"Bun" matching MCP tools:

  • Find arbitrage opportunities on Polymarket via monotonicity violations + partition-sum checks. Call with NO args for a `trending_scan` of the top ~200 markets by weekly volume; pass `event` for the strongest per-event partition_check, or `topic` for a themed cross-event scan. `event` (recommended for a specific market): pass a Polymarket event slug like "fed-decision-may-2026" or "when-will-bitcoin-hit-150k"; walks child markets, checks date-axis / threshold-axis ordering AND computes the partition_check (sum of YES prices across mutually-exclusive legs — should ≈1; deviations >3pp emit a BUY/SELL EVERY LEG signal). `topic` (for cross-event scanning): pass a seed question like "Strait of Hormuz traffic returns to normal" or "Fed rate decision"; searches related events across the platform, flattens markets, runs the comparator on the union. Cross-event mode catches "...by May 31" vs "...by Jun 30" patterns that single-event misses. SEMANTIC ANCHOR: cross-event pairs require ≥0.30 Jaccard similarity on question tokens (prevents Powell-Fed-Pause being paired with Powell-DOJ-probe); skipped_low_similarity surfaces the rejected pair count. PARTITION FILTER: drops will-person-X / will-manager-Y / will-someone-else- placeholder slugs; partitions with >20% placeholder fraction return null arb signal. Response: opportunities[] (gap_pp, suggested_trade, reasoning, monotonicity violation context), and in event mode partition_check{sum_yes_prices, gap_from_1, placeholders_filtered, suggested_trade}. FILL CHECK: when the partition signal fires, arbitrage.fill_check prices it against live CLOB depth (theoretical_edge_pp_at_book vs realizable_edge_pp at 1000 shares/leg, thin_legs[]) — realizable_edge_pp ≤ 0 means the overround exists only at last-trade, not in the book; do not trade it. For custom sizing use polymarket_fill_risk.
    Connector
  • What can I ask Pipeworx? / what is Pipeworx good for? / what can you do? / give me ideas / show me examples / getting started / what data do you have? — the onboarding entry point for an agent that just connected and wants to know what is worth asking. Returns category-bucketed example questions (company financials, drugs & clinical trials, economics, real estate, prediction markets, weather, government & patents, science & academia, news) — each with the exact tool + argument shape that answers it, drawn from the live catalog of thousands of tools. Call with no arguments for the full spread, or pass `topic` (e.g. "finance", "pharma", "betting") to focus. Use this FIRST when you do not yet know what Pipeworx can do for you, or to learn how to call the meta-tools (ask_pipeworx, entity_profile, compare_entities, etc.).
    Connector
  • ONE call to find the best small specialist model for your task, proven on YOUR examples. `task` = plain-language description (e.g. "multilingual sentence embeddings for FAQ search", or "extract product fields from scraped HTML"). Set `kind` to steer the bake-off family: embedding/rerank (retrieval) or generative/extraction (HTML->JSON). `my_examples` shape follows kind: retrieval => {query, positive, negatives:[...]}; extraction => {input, expected:{field:value}, schema?:{...}}. Chains discover->eval: searches the 147K-model catalog, bakes the top candidates off on your examples, and returns the WINNER model id + scorecard + a ready-to-call serving block (embeddings, rerank, or /v1/chat/completions for extractors). Use this as your default entry point; drop to `discover_models`/`eval_models` only to browse or hand-pick. Returns: {winner, scorecard, serve:{endpoint, model, how_to_call}}.
    Connector
  • Find tools by describing the data or task. Use when you need to browse, search, look up, or discover what tools exist for: SEC filings, financials, revenue, profit, FDA drugs, adverse events, FRED economic data, Census demographics, BLS jobs/unemployment/inflation, ATTOM real estate, ClinicalTrials, USPTO patents, weather, news, crypto, stocks. Returns the top-N most relevant tools with names, descriptions, and full input schemas (with curated examples) — each result is ready to call directly, no second schema lookup needed. Call this FIRST when you have many tools available and want to see the option set (not just one answer).
    Connector
  • Realizable-vs-theoretical edge check against live CLOB order-book depth. REQUIRES one of `market` (single-market mode) or `event` (basket/partition mode). SINGLE-MARKET: pass a market slug/URL + side (buy_yes|sell_yes|buy_no|sell_no, default buy_yes) + size_usd (default 1000 — max spend on buys, target proceeds on sells); walks the ladder and returns top_of_book, vwap_fill_price, slippage_pp, shares_filled, max_fillable_usd, and a verdict (clean|degraded|cannot_fill). BASKET: pass an event slug/URL + side (sell_yes = capture overround by selling every leg, buy_yes = capture underround; default auto from partition sum) + size_usd interpreted as settlement notional S (shares per leg; each share pays $1); returns theoretical_sum vs realizable_sum (top-of-book vs VWAP across all legs), capture_ratio, profit_usd at executed size, per-leg fill detail, thin_legs[], max_clean_notional_usd, and forced_directional_risk naming the legs most likely to strand you unhedged. USE THIS before acting on any polymarket_arbitrage SELL/BUY-EVERY-LEG signal or any polymarket_edges trade above ~$500 — theoretical overround on thin books is not capturable, and partial basket fills convert an arb into an unhedged directional position (the dominant loss mode in real arb-bot P&L).
    Connector
  • Find arbitrage opportunities on Polymarket via monotonicity violations + partition-sum checks. Call with NO args for a `trending_scan` of the top ~200 markets by weekly volume; pass `event` for the strongest per-event partition_check, or `topic` for a themed cross-event scan. `event` (recommended for a specific market): pass a Polymarket event slug like "fed-decision-may-2026" or "when-will-bitcoin-hit-150k"; walks child markets, checks date-axis / threshold-axis ordering AND computes the partition_check (sum of YES prices across mutually-exclusive legs — should ≈1; deviations >3pp emit a BUY/SELL EVERY LEG signal). `topic` (for cross-event scanning): pass a seed question like "Strait of Hormuz traffic returns to normal" or "Fed rate decision"; searches related events across the platform, flattens markets, runs the comparator on the union. Cross-event mode catches "...by May 31" vs "...by Jun 30" patterns that single-event misses. SEMANTIC ANCHOR: cross-event pairs require ≥0.30 Jaccard similarity on question tokens (prevents Powell-Fed-Pause being paired with Powell-DOJ-probe); skipped_low_similarity surfaces the rejected pair count. PARTITION FILTER: drops will-person-X / will-manager-Y / will-someone-else- placeholder slugs; partitions with >20% placeholder fraction return null arb signal. Response: opportunities[] (gap_pp, suggested_trade, reasoning, monotonicity violation context), and in event mode partition_check{sum_yes_prices, gap_from_1, placeholders_filtered, suggested_trade}. FILL CHECK: when the partition signal fires, arbitrage.fill_check prices it against live CLOB depth (theoretical_edge_pp_at_book vs realizable_edge_pp at 1000 shares/leg, thin_legs[]) — realizable_edge_pp ≤ 0 means the overround exists only at last-trade, not in the book; do not trade it. For custom sizing use polymarket_fill_risk.
    Connector

Matching MCP Servers

  • A
    license
    -
    quality
    B
    maintenance
    A lightweight HTTP-based MCP server built with Bun, enabling tool discovery and execution via JSON-RPC 2.0 over HTTP.
    15
    MIT
  • A
    license
    -
    quality
    D
    maintenance
    A modern MCP server template built with Bun and TypeScript, supporting stdio and HTTP transports for easy integration with AI assistants like Claude.
    16
    1
    MIT

Matching MCP Connectors

  • Global stock research, ML forecasts, valuation signals, screeners & portfolio tracking in Claude

  • ParlayOAuth

    Read-only MCP server for live Polymarket, Kalshi, Limitless odds; Manifold sentiment.

  • Shop and check out, in natural language, across the merchants the user has linked (DoorDash, etc.). Pass the whole ask as `request` — e.g. "order a caesar salad from Zuni on DoorDash" — and this tool runs the shopping flow for you. It is CONVERSATIONAL: this tool RETURNS a `conversation_id`; pass that SAME `conversation_id` back on every follow-up (your reply to a question, "add a coke", "yes, check out") so it continues the SAME order. Omit it (or set new_order=true) only to start a fresh order. It will ask for the delivery address and have you confirm the cart and total. CHECKOUT (which charges a one-time card) happens ONLY after the user explicitly confirms in a later message — relay the confirmation through `request` ("yes, place the order") on the SAME conversation_id. If it hands out a merchant login link (hosted connect), just reply on the SAME conversation_id once the user finishes (e.g. "done — I logged in") and it verifies the link itself. Logins started here have no pending_id, so the buy_connect / buy_connect_status pair does not apply to them. Call get_instructions FIRST for the current usage guide before your first buy.
    Connector
  • **Executes the task on the real websites** (the search, the price check, the availability lookup, the configurator, the booking flow) and returns what came back. Runs a script you authored against the `get_library` vocabulary, on the live sites, and returns `{ ok, result, logs, error, ms }`. Call `get_library` FIRST — it gives the exact function names, argument shapes, and return types; this description is the LANGUAGE + how-to (get_library is just the vocabulary). THE LANGUAGE — plain async JavaScript: • `bowmark` is a ready global (no import). Call capabilities off it — `await bowmark.<capability>.<method>(...)` — always `await`, they're async. • Individual sites are callable too, at `await bowmark.providers.<provider>.<fn>(...)`. Use one when you specifically want THAT site; otherwise prefer the capability, which fans out across sites and routes around failures. • Real control flow: `await`, `if`, loops, array methods (`map`/`filter`/`sort`/`slice`), and `Promise.all` for fan-out. • `return` a value to get it back (JSON-serialized). `log(...)` for progress lines. • `bowmark` is the ONLY I/O — no `fetch`, `process`, filesystem, or `import`/`require`. Write a plain async body, not a wrapping function. • Keep scripts small and deterministic — no infinite loops. Runs in a hard sandbox with CPU + memory + wall-clock limits. SENDING IT: pass the script text as `run({ script })` — `script` is the only argument (there is no `site` argument; the library exposes every capability under `bowmark`). `result` is whatever you returned; `logs` are your `log()` lines in order; on a throw/timeout `ok:false` and `error` is set. CHECK `status` BEFORE `ok`. It is `ok` | `error` | `partial` | `needs_user`. • `partial` means the script RAN and `result` is real and usable, but some of what it called never answered — so the result is narrower than what you asked for. `ok` is still `true`; this is not a failure. `incomplete.summary` says what happened in one sentence, `incomplete.failures` names each call that threw and what the site said, and `incomplete.degraded` names each call that answered while reporting its OWN results thin. You MUST say so when you present the result: name what was missed, and do not describe it as complete, exhaustive, or 'all' of anything. A `partial` you report as whole is a wrong answer, not a slightly smaller right one. • Before you conclude a `partial` is final, check `incomplete.failures[].fixable`. `fixable: true` means YOUR ARGUMENT was rejected, not the site — the error text names what that function actually takes, so re-read it in `get_library`, fix the argument and run again; that recovers the whole answer. For any other failure re-running usually returns the same thing. • `needs_user` means a site needs the USER signed in — it is NOT a failure and NOT something you can fix by editing the script. `needs` lists the sites; `meta.handoff.url` is a single-use link that expires (`meta.handoff.expiresAt`). Give the user that URL, say which sites it covers, and WAIT. When they tell you they're done, send the SAME script again unchanged. Do NOT retry before then — it will stop at the same place and cost another run. Do NOT try to log in yourself, ask them for a password, or work around it with a different site. • Logged-in runs need a Bowmark API key on the connection; if you get `needs_user` saying so, tell the user to add one rather than retrying. `trace` is the execution trace — every capability you called and the providers it fanned out to under the hood: `[{ kind:'capability', capability:'flights', method:'search', ms }, { kind:'provider', capability:'flights', provider:'google_flights', fn:'search', results, status, ms }, …]`. The script never visits websites — it calls capabilities that route to providers, and the trace is the receipt. Composition is the point — call a method MULTIPLE times and combine results. To sweep a date range, call the search per date inside `Promise.all` and sort/filter the merged array (each flight result carries its `date`, so you can tell the runs apart). See the `get_library` examples for the exact shape. SOME capabilities return their rows alongside a `warnings` array — `{ flights, warnings }`, `{ hotels, warnings }`, `{ cars, warnings }`. Others return a bare array. The signature in `get_library` tells you which; go by it rather than assuming. Where there IS a `warnings` array it names any site dropped from the fan-out, and the rows themselves look identical with or without it. Read it, and pass on anything it says rather than quoting a 'cheapest' that only ranks the sites that happened to answer. Dropping `warnings` from what you return does not hide it — the run comes back `status: 'partial'` regardless, because the runtime counts what your script CALLED, not what it chose to report.
    Connector
  • ACCOUNT REQUIRED (free — sign in via GitHub at https://pipeworx.io/signup; depth:"thorough" needs a paid plan). If you are not signed in, use ask_pipeworx instead — it works on every tier. Grounded multi-source research across Pipeworx's 1419 STRUCTURED data sources (SEC filings, FRED/BLS economics, FDA, USPTO patents, markets, science, government records, etc.) in ONE call — this is NOT open-web search. Decomposes your question into focused facets, routes each to the right one of 5,462 tools IN PARALLEL, and returns a findings packet: verbatim evidence + confidence + source + fetched_at + a stable pipeworx:// citation per finding, with explicit gaps[] for facets the data couldn't answer (never invented). Best for broad/multi-part questions over structured data ("compare X and Y's regulatory + financial exposure", "research the filings + market picture for ACME"). For a single lookup use ask_pipeworx (one LLM call, not many). For BREAKING or colloquial CURRENT-NEWS / "what's the world saying about X" topics, prefer ask_pipeworx — it routes to live news APIs and the *-news-feeds packs; deep_research returns mostly empty gaps[] when the topic isn't in the structured catalog. Second-hop iteration: depth:"standard" re-angles unanswered gaps (gap recovery); depth:"thorough" additionally chases the best leads from the first pass — so multi-step questions resolve in one call. Every finding carries a `hop` field and a citation_uri — a resolvable pipeworx:// record URI, present only when the source emits one that resources/read can actually serve, so a citation you get back is always fetchable. "standard" and "thorough" also return contradictions[] flagging findings that disagree. Large records are semantically excerpted to the passages relevant to each facet (not head-truncated), so answers deep in a long filing/series aren't missed. Expect 15-60s (thorough with its follow-up + contradiction pass: up to ~90s).
    Connector
  • "Tell me about X" / "research Acme" / "brief me on Tesla" / "what does Apple do" / "company profile for Microsoft" / "give me the rundown on NVDA" / "everything you know about $TICKER" — full cross-source profile of a US public company in ONE parallel call. ALWAYS PREFER over chaining single-pack SEC/XBRL/news lookups when the user asks for a holistic view. Fans out across SEC EDGAR, XBRL, USPTO, news, GLEIF and returns: cik + company_name; recent_filings (up to 5 with pipeworx://edgar/company/{cik}/filings/{accession} URIs); fundamentals (LATEST 10-K Revenues + NetIncomeLoss + Cash, sorted period_end DESC); patents (USPTO PatentsView API sunset May 2025 — soft-fails until reactivated); recent news mentions via GDELT→GNews fallback; LEI via GLEIF. Pass ticker "AAPL" or zero-padded CIK "0000320193" — names not supported (use resolve_entity first if you only have a name).
    Connector
  • ACCOUNT REQUIRED (free — sign in via GitHub at https://pipeworx.io/signup; depth:"thorough" needs a paid plan). If you are not signed in, use ask_pipeworx instead — it works on every tier. Grounded multi-source research across Pipeworx's 1419 STRUCTURED data sources (SEC filings, FRED/BLS economics, FDA, USPTO patents, markets, science, government records, etc.) in ONE call — this is NOT open-web search. Decomposes your question into focused facets, routes each to the right one of 5,462 tools IN PARALLEL, and returns a findings packet: verbatim evidence + confidence + source + fetched_at + a stable pipeworx:// citation per finding, with explicit gaps[] for facets the data couldn't answer (never invented). Best for broad/multi-part questions over structured data ("compare X and Y's regulatory + financial exposure", "research the filings + market picture for ACME"). For a single lookup use ask_pipeworx (one LLM call, not many). For BREAKING or colloquial CURRENT-NEWS / "what's the world saying about X" topics, prefer ask_pipeworx — it routes to live news APIs and the *-news-feeds packs; deep_research returns mostly empty gaps[] when the topic isn't in the structured catalog. Second-hop iteration: depth:"standard" re-angles unanswered gaps (gap recovery); depth:"thorough" additionally chases the best leads from the first pass — so multi-step questions resolve in one call. Every finding carries a `hop` field and a citation_uri — a resolvable pipeworx:// record URI, present only when the source emits one that resources/read can actually serve, so a citation you get back is always fetchable. "standard" and "thorough" also return contradictions[] flagging findings that disagree. Large records are semantically excerpted to the passages relevant to each facet (not head-truncated), so answers deep in a long filing/series aren't missed. Expect 15-60s (thorough with its follow-up + contradiction pass: up to ~90s).
    Connector
  • "What's new with X" / "latest on Y" / "what happened to Z this week / month / quarter" / "updates on Acme" / "news on Tesla recently" / "what's happening with Apple" — change feed for a company in the last N days/weeks/months in ONE parallel call. Fans out to SEC EDGAR (filings since `since`), GDELT→GNews fallback (news mentions in window — GDELT preferred, GNews when rate-limited or 5xx), USPTO (patents granted; PatentsView API sunset May 2025 so this soft-fails until reactivated). `since` accepts ISO date ("2026-04-01") or relative shorthand ("7d", "30d", "3m", "1y"). Returns structured changes[] grouped by source + total_changes count + pipeworx:// citation URIs. Use entity_profile instead when you want the static profile (filings + fundamentals + LEI + patents) regardless of window.
    Connector
  • Composite "should I add this npm package to my project" check in ONE call — fans out across deps.dev (license + advisories + version history) and bundlephobia (gzipped/minified bundle size, dependency count, ESM/tree-shake support). Use whenever an agent asks "is X safe / popular / small" or "what does adding lodash cost me". Returns a summary block (is_latest, license, published_at, advisory_count, bundle_kb_min, bundle_kb_gz, dependency_count, has_esm, tree_shakeable), per-advisory detail, links, and a list of recent alternative versions. NPM ecosystem only in v1; PyPI / Maven / Cargo / Go fall under deps.dev:version directly. Partial failures degrade gracefully — bundlephobia's first measurement on a new version can take 5-30s; sources_failed will list it if it times out, the rest still returns.
    Connector
  • Realizable-vs-theoretical edge check against live CLOB order-book depth. REQUIRES one of `market` (single-market mode) or `event` (basket/partition mode). SINGLE-MARKET: pass a market slug/URL + side (buy_yes|sell_yes|buy_no|sell_no, default buy_yes) + size_usd (default 1000 — max spend on buys, target proceeds on sells); walks the ladder and returns top_of_book, vwap_fill_price, slippage_pp, shares_filled, max_fillable_usd, and a verdict (clean|degraded|cannot_fill). BASKET: pass an event slug/URL + side (sell_yes = capture overround by selling every leg, buy_yes = capture underround; default auto from partition sum) + size_usd interpreted as settlement notional S (shares per leg; each share pays $1); returns theoretical_sum vs realizable_sum (top-of-book vs VWAP across all legs), capture_ratio, profit_usd at executed size, per-leg fill detail, thin_legs[], max_clean_notional_usd, and forced_directional_risk naming the legs most likely to strand you unhedged. USE THIS before acting on any polymarket_arbitrage SELL/BUY-EVERY-LEG signal or any polymarket_edges trade above ~$500 — theoretical overround on thin books is not capturable, and partial basket fills convert an arb into an unhedged directional position (the dominant loss mode in real arb-bot P&L).
    Connector
  • Verify the cryptographic integrity of an evidence bundle (ev_...) owned by your API key. Checks manifest hash, EIP-191 signature, and R2 artifact hashes. Free — no credits consumed. Use when you need to confirm a bundle has not been tampered with. For quick metadata lookups (without full crypto verification), use bundle.get instead. Also returns a signed action receipt (rcpt_...) binding this verify call to the bundle manifest — list with receipt.list, verify with receipt.verify. Returns: { valid: boolean, bundle_id, manifest_sha256, checks: { status, manifest_hash, signature, artifacts: [{ name, ok }] }, tampered: string[], signer_address: string|null, attestation_tx: string|null, url: string, captured_at: string, receipt: ActionReceipt|null } Example prompts: - "Verify the cryptographic integrity of bundle ev_550e8400." - "Is this evidence bundle still valid and untampered?" - "Deep-check the manifest hash and signature of my bundle."
    Connector
  • Verify the cryptographic integrity of an evidence bundle (ev_...) owned by your API key. Checks manifest hash, EIP-191 signature, and R2 artifact hashes. Free — no credits consumed. Use when you need to confirm a bundle has not been tampered with. For quick metadata lookups (without full crypto verification), use bundle.get instead. Also returns a signed action receipt (rcpt_...) binding this verify call to the bundle manifest — list with receipt.list, verify with receipt.verify. Returns: { valid: boolean, bundle_id, manifest_sha256, checks: { status, manifest_hash, signature, artifacts: [{ name, ok }] }, tampered: string[], signer_address: string|null, attestation_tx: string|null, url: string, captured_at: string, receipt: ActionReceipt|null } Example prompts: - "Verify the cryptographic integrity of bundle ev_550e8400." - "Is this evidence bundle still valid and untampered?" - "Deep-check the manifest hash and signature of my bundle."
    Connector
  • Notarize an evidence bundle on-chain by writing its manifest SHA-256 to the blockchain (Base/EVM). Creates a permanent, tamper-evident on-chain record of the document fingerprint. If the bundle is already notarized, returns the existing attestation immediately (idempotent). Use when you need an immutable on-chain timestamp proving a document existed at a point in time. For quick integrity checks without on-chain cost, use bundle.verify instead. Also returns a signed action receipt (rcpt_...) binding this notarize call to the bundle manifest — list with receipt.list, verify with receipt.verify. PREREQUISITE: Bundle status must be "complete". Check status with bundle.get first. NOTE: Costs gas (ETH). The on-chain record is permanent and cannot be deleted even if the bundle is later purged. Returns: { bundle_id, attestation: { tx_hash, network, attested_at, key_id, eas_uid?, schema_uid? }, receipt: ActionReceipt|null } Example prompts: - "Notarize bundle ev_550e8400 on-chain so I have a permanent record." - "Put the fingerprint of my evidence bundle on the blockchain." - "Create an on-chain timestamp for this document bundle."
    Connector
  • Pay once, then call without per-request settlement — Buy a prepaid balance in a single x402 settlement and get a secret credit token. Send it as the `x-credit-token` header on any paid service and each call debits its price from your balance — no per-call signature, no settlement latency. Built for agents that fire many checks a minute. Tiers: $0.25 (starter), $1, $5 (+10%), $20 (+20%). The token is shown once; balance lasts 180 days. Priced $5.00 per call over x402 on Base; send a prepaid x-credit-token header for unlimited calls, or get 1 free call/day per tool. No wallet or API key required.
    Connector
  • PAID PER DELIVERY (30 credits each time it fires, NOT at creation — free to create/cancel) — the only MCP tool that exposes GISGP's core recurring-export product to agents: creates a schedule that re-exports a FeatureServer layer on its own and POSTs the file straight to your own webhook_url, no email/web UI account needed beyond the API key. Reuses the same scheduler that runs the paid web app's scheduled exports (fires within ~5 min of the due time). format: "csv", "geojson", "shapefile", "kml", or "excel". frequency: "hourly" (top of each hour), "daily" (at run_hour UTC), "weekly" (at run_hour UTC on `weekday`, 0=Monday..6=Sunday), or "monthly" (at run_hour UTC on `monthday`, 1-28). webhook_url: must be a public, reachable HTTPS URL (validated at creation AND at every delivery) — GISGP POSTs a JSON body {schedule_id, format, row_count, filename, delivered_at, data_base64 (or download_url for files >5MB)}. If the wallet lacks 30 credits when a delivery is due, that cycle is silently skipped (schedule stays active, no error surfaced to you) — top up any time at https://gisgp.com/billing/mcp-credits/topup and the next cycle delivers normally. Use estimate_cost or check_wallet_balance to plan ahead. Delete via cancel_export_schedule when no longer needed — an abandoned schedule with an empty wallet just skips forever, but does not charge or error. Requires Authorization: Bearer <api_key>. Returns JSON: {ok, schedule_id, next_run_at}.
    Connector
  • Prove which candidate model actually wins on YOUR data before committing — a head-to-head bake-off, not a public leaderboard (MTEB rank does NOT predict your-domain performance). `task` = 'embedding'|'rerank' (rank a labeled positive), 'classification' (score a classifier's top-1 label vs your gold labels), OR 'extraction'|'generation' (bake off generative extractors). `models` = candidate HF ids (e.g. from `discover_models`). `dataset` shape depends on task: for embedding/rerank each item is {query, positive, negatives:[...]}; for classification each item is {text, label}; for extraction/generation each item is {input, expected:{field:value}, schema?:{...}}. Ranking tasks return MRR/hit@k/ndcg@k; classification returns accuracy/macro_f1; extraction returns field_accuracy / json_validity / exact_record. Always returns the single `winner` model id — feed it into `embed`/`rerank`/`classify`/`extract` to serve it. First call cold-loads each model; be patient.
    Connector