Research you can read.
Tools you can call.
Use Aguasto’s published research and educational calculations from your application or AI assistant. Public access requires no API key.
Available today
Search descriptions of our public GPU, model and infrastructure references, follow their original sources, and run the same educational formulas as our engineering tools. The platform is in development: GPU capacity is not bookable, and inference, reservations, service accounts, API keys and token pricing are unavailable.
HTTP quickstart
Send JSON requests over HTTPS. The OpenAPI description defines request fields, responses and errors. The public page catalog includes canonical references and modification dates.
curl https://www.aguasto.com/ask \
-H 'Content-Type: application/json' \
-d '{"query":{"text":"GPU rack cooling"},"prefer":{"mode":"list"}}'
curl https://www.aguasto.com/api/v1/calculate \
-H 'Content-Type: application/json' \
-d '{"calculation":"weight_memory","parameters":{"parameters_billions":70,"bits":4}}'Search uses deterministic matching of published titles and descriptions. It returns references, rather than generated answers. Follow each canonical page for full evidence. Request prefer.streaming: true for NLWeb 0.55 events: start, result or error, and complete.
For smaller catalog responses, request /api/v1/catalog?limit=20&offset=0. Continue with pagination.next_offset until it is null. Omitting pagination parameters returns the complete current inventory.
Read several page records in one request
Send up to 20 unique published paths to /api/v1/pages/batch. Results preserve your order and include metadata and canonical links. Unknown paths receive a per-item 404 while valid records still return. External URLs, query strings and duplicate paths are rejected. This reads public metadata without fetching arbitrary URLs or storing inputs.
curl https://www.aguasto.com/api/v1/pages/batch \
-H 'Content-Type: application/json' \
-d '{"paths":["/gpus","/models"]}'Try a public request
This playground searches public descriptions or calculates an illustrative estimate. It has no customer data, account or provisioning actions. Don’t enter confidential information. The application does not store these inputs; hosting providers may process technical request metadata.
Connect with MCP
Add a Streamable HTTP server in your MCP client using one of these URLs. Both are anonymous and read-only. Use the tools advertised by the live tools/list response.
https://www.aguasto.com/mcp/docs— search published page descriptions and retrieve canonical references.https://www.aguasto.com/mcp/tools— estimate weight memory, facility energy or completed-token cost.
{"mcpServers":{"aguasto-research":{"type":"http","url":"https://www.aguasto.com/mcp/docs"},"aguasto-calculators":{"type":"http","url":"https://www.aguasto.com/mcp/tools"}}}Client configuration formats vary. MCP requests use Accept: application/json, text/event-stream. Responses use stateless JSON; server-initiated SSE and sessions are not offered. See research server card and calculator server card.
Verified MCP Registry records: Aguasto Research and Aguasto Engineering Calculators. The registry identifiers use the verified www.aguasto.com hostname. Client connections continue to use the HTTPS endpoints above.
You can also find the public research server in the Aguasto Research listing on MCP.so, with the direct endpoint, authentication setting and client setup instructions.
Agent skills and browser tools
Read the research skill or calculation skill. The plugin manifest and MCP configuration describe the package. The official public integration repository contains standalone clients, skills, plugin configuration and usage rules. It excludes the private corporate website source and retains the existing rights-reserved terms. Install the verified public skills with npx skills add Altssdata/aguasto-agents --skill aguasto-research aguasto-calculations. Directory pages: research and calculations.
In browsers that implement WebMCP, the site registers public research search and weight-memory calculation tools through document.modelContext, with a navigator.modelContext compatibility fallback. Ordinary navigation and calculators work when this experimental API is absent.
Install the Go client or Homebrew CLI
The official Go SDK uses the standard library and supports public catalog, search and calculation requests. The Homebrew formula installs the SHA-256-pinned JavaScript release with Node.js. Both expose the same public research interfaces.
go install github.com/Altssdata/aguasto-agents/go/cmd/aguasto-research@v1.0.0
brew tap altssdata/aguasto-agents https://github.com/Altssdata/aguasto-agents
brew install altssdata/aguasto-agents/aguasto-research
aguasto-research fixtures
aguasto-research search "GPU rack cooling"JavaScript SDK and CLI on npm
Install the official aguasto-research npm package with Node.js 22 or newer. The package and CLI support public search, page references and educational calculations.
npm install aguasto-research@1.0.0
npx aguasto-research fixtures
npx aguasto-research search "GPU rack cooling"Source archives and Python client
The JavaScript SDK and CLI and Python SDK and CLI are versioned source packages for these public interfaces. The JavaScript package is also published on npm; the Python client is available as a source download and is not published on PyPI. You can install either downloaded archive locally. See the included README and use the CLI’s --help command. Never install a similarly named registry package on the assumption it is official.
To use the Aguasto command-line interface with Node.js 22 or newer, download the JavaScript archive above into your project, then run:
The same standalone SDK/CLI source archives are available in the verified v1.0.0 public integration release.
npm install ./aguasto-research-js.tgz
npx --no-install aguasto-research --help
npx --no-install aguasto-research fixtures
npx --no-install aguasto-research search "GPU rack cooling"Access, errors and safe testing
See public access instructions. No login or token exchange is required. JSON bodies are limited to 16 KiB; queries to 2,000 characters and 20 results. Malformed inputs fail with structured errors. Cross-origin browser access is restricted to the configured site origin. No credential or paid compute flow is exposed.
The playground and calculators do not change production data. For offline integration tests, the downloadable packages include fixtures; they are not a customer GPU sandbox. Search and calculation requests never call a model provider.
API versioning and deprecation policy
The standalone versioning and deprecation policy records the current contract, compatibility rules and retirement notices. Catalog, batch and calculator responses link it with Link: rel="deprecation"; that relation makes the policy discoverable without declaring an active operation deprecated.
REST operations use the /api/v1 major version. New optional fields may be added within v1; clients should ignore unknown response fields. Breaking request or response changes use a new major URL version. Before removing a deprecated operation, we will document its replacement and planned removal date here at least 90 days ahead and advertise that date with Deprecation and Sunset response headers. No v1 operation is currently deprecated.
REST errors use a stable Error schema with error.code and error.message. Invalid inputs return 400; unknown v1 endpoints return JSON with HTTP 404. NLWeb retains its own documented failure envelopes, and MCP retains JSON-RPC errors.
Retrying calculations with a content-bound key
Calculator POST requests evaluate pure v1 functions and do not change server data. Retrying the same inputs produces the same result across instances. For an explicit input-consistency check, send an optional Idempotency-Key structured string containing "sha256:<digest>". The digest is SHA-256 of the compact UTF-8 JSON envelope, with root keys calculation then parameters and parameter keys in lexicographic order, using JavaScript JSON number/string encoding. For example:
import { createHash } from "node:crypto";
const payload = JSON.stringify({calculation:"weight_memory",parameters:{bits:4,parameters_billions:70}});
const key = '"sha256:' + createHash("sha256").update(payload).digest("hex") + '"';
const response = await fetch("https://www.aguasto.com/api/v1/calculate", {
method:"POST", headers:{"Content-Type":"application/json","Idempotency-Key":key}, body:payload
});The server verifies the content binding and echoes a valid key. Malformed keys return 400; a key paired with different inputs returns 422. Keys have no time-based expiry within the v1 formula contract. The server recomputes these deterministic estimates; it stores neither request keys nor inputs and offers no stored-job or transaction replay service. Requests without this optional header continue to work.
REST responses expose RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset (seconds until reset), plus compatibility X-RateLimit-* fields whose reset value uses Unix seconds. Within each warm server instance, catalog, index and batch page reads share 120 requests per minute and calculations allow 30, keyed by a temporary hash of the provider-supplied client address. Exceeding a window returns 429 and Retry-After. These limits reset on cold starts and apply independently across instances; they are best-effort admission controls, not global quotas. The hosting platform may impose additional limits.
Discovery and reference data
Catalog provenance links pin the public skill sources and MCP client configuration to a reviewed GitHub revision with SHA-256 digests. This publisher metadata is unsigned advisory information; it does not assert compliance certification or cryptographically verified publisher identity.
- Agentic Resource Discovery catalog
- API catalog and skill index
- Website index and developer index
- Schema.org feed and Schema Map
- Structured agent view
For future enterprise compute requirements, use our contact instructions. Online delivery remains disabled; an inquiry draft is unsent until you send it using your own email application.