io.github.nmxmxh/pronto-stream

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Real-time planetary signal engine and Model Context Protocol (MCP) server for autonomous AI agents.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "pronto-stream": {
      "url": "https://pronto.stream/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (15)

• get_billing_profile read-only

  Report who you are resolved as and what your plan grants: the agent identity behind this call, the funded account it acts for, the plan, the per-minute call ceiling, the per-response row budget, and which indicator set you may compute. Returns a CWF BILL line, a CREDIT line when a metering balance exists, and a NOTE line naming your state and its remedy when you are not entitled — unauthenticated, or registered but not yet claimed by a funded account. If unauthenticated, call register_agent next: it creates an identity with no human present. From there you can either buy your own plan with open_payment, or hand the claim code it returns to whoever holds your budget.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• register_agent

  Create your own identity on pronto and receive an API key, with no human present. Send the key back as an Authorization header on later calls and you will be recognised across sessions. Registering grants no plan. A new identity has exactly the limits an unauthenticated caller has. There are two ways to raise them, and you can take either. If you hold your own funds, call get_plans and then open_payment — you buy your own plan and no person is involved at any point. If your costs belong to someone else, hand them the claim code and URL in this response; once they claim you, your calls resolve to their plan with no change on your side. Supply an Ed25519 public key (base64) if you can keep the private half: it lets you rotate your own credential before it expires without anyone's help. Without one you cannot rotate, and you must register again when the key expires.

  Parameters:

  • label — string — A name for this agent, for whoever reviews the account later.
  • public_key — string — Base64-encoded Ed25519 public key, 32 bytes. Optional, but without it you cannot rotate your own credential.

  Hints: read-only no · destructive not declared · idempotent no · open-world yes

• open_payment

  Buy access for yourself, with no human involved. Step one of two: this quotes the exact amount and binds the purchase to the wallet you will pay from. Three things can be bought. Call credits: plan "credits" with amount_usd from 1 to 100, at $0.001 a call — one credit per successful tool call, Builder's limits while the balance lasts, no expiry, failed calls not charged. A pass: plan "builder" or "team" with cycle "day" or "week". A plan period: the same plans with cycle "monthly" or "yearly". Nothing renews automatically; buying the same plan again before it ends extends it. You must already have an identity — call register_agent first and send the key as an Authorization header. Call get_plans for the ladder and what each option raises. from_address is required and is the wallet the USDC will leave. It is bound now, before the payment exists, because a transaction hash is public the moment it is mined: without the binding, anyone watching the settlement address could take your transfer and redeem it as their own. Only a transfer from the wallet you name here, redeemed with a signature from that same wallet, will settle this purchase. Returns PAY lines with the atomic amount, the token, the network and the address to pay. Transfer it, then call settle_payment with the transaction hash.

  Parameters:

  • plan (required) — string — What to buy: call credits, or a plan. Enterprise is a design partner engagement and cannot be bought here.
  • from_address (required) — string — The EVM address the USDC will be sent from, 0x followed by 40 hex characters. Binding, and required.
  • cycle — string — Period for a plan. Defaults to monthly. Ignored for credits.
  • amount_usd — integer — Whole dollars of credits to buy, at $0.001 a call. Required when plan is credits.

  Hints: read-only no · destructive not declared · idempotent yes · open-world yes

• settle_payment

  Redeem a payment you have made and activate what it bought — a plan, a pass or call credits. Step two of two, after open_payment and the transfer. Prove the payment is yours: sign the SIGN message from open_payment with the paying wallet (personal_sign; smart wallets via ERC-1271) and pass it as signature. A transaction hash is public once mined, so without the signature anyone who saw your transfer could redeem it. The payment is verified by reading Base: the transfer must have reached the settlement address, from the wallet this purchase was opened with, for at least the amount owed, and be five blocks deep. A payment that is real but not yet deep enough returns a pending error naming that — the money has moved, so do not pay again: it settles on its own once deep enough, and calling this again with the same hash reads the outcome. One transaction pays for one purchase, once. On success your existing API key resolves to the new plan with nothing to change on your side: you become your own principal, and no person holds the subscription.

  Parameters:

  • transaction_hash (required) — string — The Base transaction hash of your USDC transfer: 0x followed by 64 hex characters.
  • signature (required) — string — 65-byte hex signature (0x…) from the paying wallet, made with personal_sign (EIP-191) over the SIGN message open_payment returned, with your lower-case transaction hash filled in. Smart-contract wallets are checked with ERC-1271.

  Hints: read-only no · destructive not declared · idempotent yes · open-world yes

• get_plans read-only

  List what can be bought — call credits, day and week passes, and plans — what each costs, and what each raises — the call ceiling, the row budget, and whether the commercial indicator set is included. Machine-readable, so you can decide what to do about a refusal without parsing a pricing page: buy the plan yourself with open_payment, or escalate to whoever holds your budget. Returns a CREDITS line, one CWF PLAN line per plan with its pass and period prices, and a SETTLE line naming the asset, network and address to pay.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_latest_signals read-only

  Fetch the most recent live signals from the wire in token-efficient CWF format, optionally filtered to one collection. The all-collections view is diversified: no provider may exceed its share of the page, no collection may take more than half of it, and what overflows is named on a note line rather than dropped silently — so a burst of tickers cannot crowd out earthquakes. Pass collection to read one collection at full depth.

  Parameters:

  • limit — number — Maximum number of signals to return (default 20, max 100)
  • cursor — string — Page token from META cursor: of the previous response. Omit for the first page.
  • relate — boolean — Also return the link graph over the rows in this response, as REL and CLUSTER lines, so repeats of one story collapse into one line without a second call. Off by default.
  • collection — string — Signal collection to filter by, e.g. climate, sentiment, security, health, cosmic, energy, macro, market, crypto, knowledge, prediction, logistics. Omit for all collections.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• get_signal read-only

  Fetch one signal by its id, when you already know which one you want. Every other read tool returns a set and bounds it, so re-reading a single row through them means re-running a query and hoping the row survives the row budget a second time; this returns that row and nothing else. Takes either the full correlation id or the short handle in column 2 of any CWF row — cite the handle, pass it back here. The live window turns over in about two minutes, so a signal read several calls ago may have aged out; the error says which of the two happened rather than leaving you to look for a typo. To find signals rather than re-read one, use search_signals or scope_signals; to reach the ones about the same subject, use get_related_signals.

  Parameters:

  • correlation_id (required) — string — Id of the signal to fetch. Accepts either the full correlation id or the short handle in column 2 of any CWF v3 row.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• search_signals read-only

  Find signals by the words in them. Use this when you are looking for a subject — a storm name, a company, a phrase a report would print — and cannot name it as a category. When you can name it as a category instead (a country, a domain, a severity), scope_signals selects that slice exactly and does not depend on any particular word appearing; this tool ranks by word overlap and will miss a matching signal that phrased it differently. The search covers places, observations, summaries, identifiers, and the country and topic facets of each signal, so "Japan" reaches a Japanese-language article that never writes the word. A signal is returned when it contains the words you asked for. Inflections count: "flood" reaches "flooding" and "quake" reaches "quakes". Synonyms do not: the match is lexical, not semantic, so "car" does not reach "automobile" and "downturn" does not reach "recession". Use the words the source would have used. Results are ranked by how much of your query each signal contains, exact phrase matches first. When nothing contains your terms the tool returns an error rather than the closest available rows; an empty result means the wire does not carry it, not that the search gave up. Returns CWF lines.

  Parameters:

  • query (required) — string — Keywords describing the anomaly or topic, e.g. 'earthquake pacific' or 'solar flare'
  • sort — string — relevance (default): word coverage, exact phrase first. significance: the server's composed attention score, which lifts severity and connectivity over freshness — use it when looking for consequence rather than currency.
  • limit — number — Maximum matches to return (default 10, max 50)
  • cursor — string — Page token from META cursor: of the previous response.
  • locale — string — Optional BCP-47 language code (e.g. 'ja', 'zh', 'de', 'ar'). When set, response metadata is returned in the requested language context.
  • min_severity — string — Drop results below this published band. Unrated rows are dropped too, because absence of a rating is not evidence of safety.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• list_fusion_products read-only

  List every derived intelligence product pronto can compute, with the published index or model each one implements, the signal collections it consumes, and whether it needs a plan. Read this to find out whether pronto computes the thing you need, then call get_fused_signal to actually compute it. The two tools answer different questions and the difference matters. This one states the catalogue, reads no signals, and returns the same answer on a wire that is still warming up. get_fused_signal states what the current window can support, so a product missing from its output means the inputs are not flowing right now, not that the product does not exist. Each row states its attribution: "published" is a faithful implementation of an external, citable index, "composed" is authoritative published inputs combined by a rule that is pronto's own and disclosed in the method. Rows marked plan:team are the commercial macro and energy models; when your plan does not include them a NOTE line names them, so a gated product is never mistaken for a missing one.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_fused_signal read-only

  Compute derived intelligence products from the live wire. Each result states the method it implements and cites the signals it was computed from; a source row with no citation is labelled as such. Pick one from the `product` enum, or omit it for everything the current window supports. A product absent from the output has no live inputs right now — list_fusion_products states the whole catalogue and does not depend on the window. Cross-source products exist only because independent publishers are time-aligned on one wire; no upstream API emits them. seismic_solution_consensus is the sharpest case: several national agencies locate one earthquake and it reports how far apart their solutions are, in magnitude, depth and true three-dimensional hypocentre. Read it before acting on any single magnitude — agencies routinely differ by half a unit while an event is still being located, which is a factor of five in energy. recession_risk_12m and sahm_recession_indicator disagree on purpose: the first forecasts, the second says a downturn has probably already begun. Read both. Some indicators are included from the Team plan. A response that withheld one says so on a NOTE line naming what and why, so an absence is never mistaken for a quiet wire.

  Parameters:

  • product — string — One product to compute. Omit for every product the current window supports.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• scope_signals read-only

  Select a slice of the live wire by facet tags — the structured metadata each signal carries, not the words in its text. Use this when you can name the slice you want: every high-severity climate signal in Japan, everything from one provider, everything in one language. It returns the whole slice whatever words the signals happen to use, which is what separates it from search_signals; reach for search when you have a subject word and no category to name. Facets cover language, country, region, domain, topic, severity, provider, coverage, and place. Facets that share a key combine with OR. Facets with different keys combine with AND. The filter "country:JP,country:CL domain:climate" selects Japan or Chile, and climate. A locale filter alone cannot answer this question.

  Parameters:

  • limit — number — Maximum signals to return (default 25, max 200)
  • cursor — string — Page token from META cursor: of the previous response. Omit for the first page.
  • filter — string — Facet expression: space- or comma-separated key:value pairs. Keys are fixed vocabulary — lang, country, region, domain, topic, severity, provider, coverage, place. Same key = OR, different keys = AND. Examples: "country:jp domain:climate" (Japanese climate signals), "country:us,country:cl severity:high" (US or Chile, high severity), "place:Tokyo" (resolved to a gazetteer id). Call list_facets for live values.
  • locale — string — Optional BCP-47 language code (e.g. 'ja', 'zh', 'de', 'ar'). When set, response metadata is returned in the requested language context.
  • relate — boolean — Also return the link graph over the rows in this response, as REL lines, so repeats of one story can be collapsed without a second call. Off by default.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• get_facet_manifest read-only

  Read the shape of the wire in one small response: a two-level tree of domain and the topics under it, each with a signal count, ranked so the busiest lead. This is the map to read first — read it once, choose a scope offline, then make one precise call to scope_signals or search_signals instead of guessing a filter. It covers two facet keys only, domain and topic, and truncates the topic tail under each domain, so it stays short enough to read in full. When you need the rest of the vocabulary — the languages, countries, providers, severities, coverages and place ids you can also filter by, exhaustively and with counts — call list_facets instead. The manifest states structure, not signal content.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• list_facets read-only

  List the full filter vocabulary: every facet value currently on the wire, across every key — lang, country, region, domain, topic, severity, provider, coverage and place — as one flat list with a signal count each, and a human-readable label for place ids. Every value it returns is a value scope_signals will accept right now. This is the exhaustive listing and it is long, running to a hundred-odd values on a busy wire. If all you need is to pick a domain and a topic, get_facet_manifest answers that in a fraction of the tokens; come here when you need a value the manifest does not carry, or the counts behind one. Place facets are gazetteer ids; filter them by name with scope_signals rather than by the raw id.

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• get_related_signals read-only

  Follow the link graph from one signal to other signals about the same subject. Each link states the reason the link exists. An echo link marks the same story from several sources. A related link marks a similar signal in the same domain. A cross-domain link joins two domains. For example, a seismic record and a news headline name the same place. The two signals arrive from unrelated providers and meet only on the wire.

  Parameters:

  • correlation_id (required) — string — Correlation id of the signal to start from. Accepts either the full id or the short handle in column 2 of any CWF v3 row. Aliases accepted for robustness: id, signal_id.
  • limit — number — Maximum neighbours (default 5, max 10)

  Hints: read-only yes · destructive no · idempotent yes · open-world yes

• list_data_sources read-only

  List the data-source endpoints feeding the wire, with provider, signal type, country, cadence, and license. Returns one SRC line per source.

  Parameters:

  • signal_type — string — Filter by signal type, e.g. climate, security, health. Omit for all sources.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

Repository

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.