Unhosted AI Developer Portal

Build on Unhosted AI: the Intelligence API (derived token, trader, wallet and market intelligence behind an API key), the copy-trading SDK for wallets and terminals, the OpenAPI specification, and the docs. Self-serve API keys, sandbox included.

The Intelligence API serves derived intelligence — scores, gates, verdicts and one-sentence rationales computed by our own engine over live on-chain data — never raw provider rows. Base URL https://api.unhosted.ai/v1 (canonical) or https://www.unhosted.ai/api/v1 (legacy, same deployment). Every endpoint is described in the OpenAPI spec and the JSON catalog, and every error is JSON.

  • OpenAPI 3.0 spec — all operations with unique operationIds, typed parameters and response schemas — import it into Swagger UI, Postman or an LLM tool-calling runtime
  • Endpoint catalog (JSON) — public, no key: every endpoint with inputs, outputs and unit cost
  • API catalog (RFC 9727) — machine-readable index of our APIs (application/linkset+json)
  • Agent Skills index — the Agent Skills we publish — drop-in instructions your agent can load for the Intelligence API and copy trading
  • MCP server — the public MCP server at https://www.unhosted.ai/mcp (no key): search and read the AI price-prediction reports from any MCP client; this link is its Server Card
  • ARD capability manifest — every machine-readable resource on this domain in one ai-catalog manifest, typed by media type — what an agent registry ingests
  • Get a key — self-serve partner console: sandbox org, API keys, usage, webhooks
  • Documentation — product, wallet SDK and copy-trading docs (Markdown versions at docs.unhosted.com/llm/…)
  • llms.txt — site overview for AI agents, including when to use Unhosted

Get an API key

  1. Sign in at the partner console with a wallet or Google/email — no invite, no approval step.
  2. Create your sandbox org, open Keys and mint a key. Sandbox keys start with key_test_ and run against real leader data in paper mode: real gates, real direction and notional, executions recorded as skipped/paper, structurally unable to reach mainnet.
  3. Go live when your flows are proven: the org owner opens Settings → Go live, reads the pricing and terms and accepts. Every key_test_ key is revoked at that moment — mint key_live_ keys from the Keys panel and redeploy with those.

Keys are hashed at rest and shown exactly once. Revocation takes effect within the key cache TTL (seconds, not a deploy). For white-glove onboarding or questions, book an integration call or email hi@unhosted.ai.

Authentication, plans and rate limits

Send the key as Authorization: Bearer <key> or x-api-key: <key>. Execution routes (/v1/copy/agent/*) additionally carry the end user's wallet authorization — X-Wallet-Address, X-Wallet-Signature (EIP-191 personal_sign of the canonical agent-API message) and X-Auth-Timestamp (valid for 300 s) — because two parties are involved: the partner who is billed and the user whose funds move. The copytrading SDK builds those headers for you (createWalletAuthHeaders).

PlanRequests per minuteNotes
sandbox60paper mode, key_test_ keys
free30evaluation
pro300production
enterprise2000custom contracts

Limits are enforced per key and reported on every response as X-RateLimit-Limit / X-RateLimit-Remaining; exceeding them returns 429 {"error":"rate_limited"}. A global daily data budget protects the shared provider quota: when it is exhausted the API returns 503 {"error":"capacity"} with Retry-After.

Metering and usage

Calls are billed in weighted units, not requests: a market-regime read costs 1 unit, a smart-money radar crawl 8, the whole token dossier (/v1/token-overview) 19. The cost of every endpoint is in the catalog (cost) and in the OpenAPI spec (x-cost-units); each response carries X-Usage-Units. Only successful calls are billed. GET /v1/usage returns your own units billed today and the service's remaining data budget — it is never rate limited, so you can always read the meter.

Quickstart

Market regime (GET, 1 unit) and a full token profile (POST, 1 unit):

export UNHOSTED_API_KEY=key_test_...

curl -s "https://api.unhosted.ai/v1/market-regime?chain=eth" \
  -H "Authorization: Bearer $UNHOSTED_API_KEY"

curl -s -X POST "https://api.unhosted.ai/v1/token-intel" \
  -H "Authorization: Bearer $UNHOSTED_API_KEY" \
  -H "content-type: application/json" \
  -d '{"chain":"eth","token":"0x6982508145454ce325ddbe47a25d4ec3d2311933"}'

Responses are derived objects — bands, scores, verdicts — e.g. { "verdict": "...", "risk": { ... }, "momentum": { ... }, "entry": { ... } }. The verdict vocabulary is fixed and load-bearing across the API: mirror-worthy, watch, avoid, unrated — and unrated means unknown, never bad.

  • Token: /v1/token-intel, /v1/token-risk, /v1/token-signal, /v1/token-holders-insight, /v1/token-narrative, /v1/token-price-action, /v1/smart-money-flow, /v1/exit-risk, /v1/token-overview (the whole dossier in one call).
  • Market: /v1/market-regime, /v1/market-movers, /v1/robinhood-movers, /v1/smart-money-radar, /v1/conviction-board, /v1/trader-board.
  • Trader and wallet: /v1/trader-score, /v1/trader-style, /v1/trader-edge, /v1/trader-leaderboard, /v1/wallet-health, /v1/wallet-analyze, /v1/wallet-pnl-quality, /v1/wallet-trades, /v1/portfolio-risk, /v1/wallet-overview (the whole trader page in one call).
  • Copy trading: /v1/copy-decision, /v1/copy-plan, /v1/copy-simulate, /v1/copy-overlap, /v1/perp-wallet, /v1/prediction-wallet, /v1/copy/top-traders, /v1/copy/spot-leaders, /v1/copy/perp-leaders, /v1/copy/prediction-leaders, /v1/copy/strategies, execution relay /v1/copy/agent/*.
  • Track records: /v1/signals/track-record, /v1/traders/track-record, /v1/perp-track-record, /v1/prediction-track-record — did the grades carry signal? Misses included.

Errors

Every error is JSON with a stable machine-readable code and a human sentence: { "error": "invalid_api_key", "detail": "Provide a valid key via Authorization: Bearer or x-api-key." }. Codes: invalid_api_key (401), wallet_auth_required / wallet_auth_expired (401), invalid_address, invalid_chain, unknown_param (400), venue_not_entitled (403), not_found / not_a_token (404), rate_limited (429), capacity, gateway_unconfigured, execution_unavailable (503), internal (500). Unknown API paths return 404 {"error":"not_found", "hint": ..., "docs": ...} rather than an HTML page. The full Error schema is in the OpenAPI spec.

Copy-trading SDK

Embeddable, non-custodial copy trading for wallets, dapps and exchanges: your users pick a graded leader, set risk limits, sign once (perps) or send one transaction (spot), and the backend mirrors the leader's trades within those limits. Venues: EVM spot via ERC-7579 smart sessions (spend cap, hard expiry), Hyperliquid perps via revocable agent keys, Polymarket predictions. The backend can submit trades and can never withdraw.

npm install copytrading

For AI agents

  • llms.txt describes the site, says when Unhosted is the right tool, and links the Markdown version of every key page.
  • Key pages negotiate content: request them with Accept: text/markdown (RFC 9110) and you get text/markdown; charset=utf-8 with Vary: Accept; or fetch the .md sibling directly (/developers.md, /index.md). Unknown URLs return a Markdown 404 that lists where to look next.
  • Unknown API paths return JSON errors; every Intelligence endpoint has a unique OpenAPI operationId, typed inputs and a response schema, so the spec can be loaded as a tool definition as-is.
  • /.well-known/api-catalog (RFC 9727) points at the spec and the docs.
  • /.well-known/agent-skills/index.json lists the Agent Skills we publish (discovery RFC v0.2.0): unhosted-intelligence-api for calling the API, unhosted-copy-trading for embedding copy trading. Each entry carries the SKILL.md URL and its SHA-256 digest, so a client can verify what it loaded and skip re-fetching an unchanged skill.
  • /.well-known/ai-catalog.json is the ARD capability manifest an agent registry ingests: one typed entry per machine-readable resource on this domain — the MCP server card, the OpenAPI document, both Agent Skills, the endpoint catalog, llms.txt, the integration guide and the research datasets — each with a stable urn:air: identifier and the natural-language queries it answers. Served with Access-Control-Allow-Origin: *.
  • The public MCP server at https://www.unhosted.ai/mcp (Streamable HTTP, no key) has two read-only tools: find_token_predictions searches the AI price-prediction reports by name, ticker or contract address, and get_token_prediction reads one as a short brief with its report link. It speaks the stateless 2026-07-28 revision and the 2025 ones. Its Server Card is also served at /.well-known/mcp/server-card.json. To connect Claude Code: claude mcp add --transport http unhosted https://www.unhosted.ai/mcp.
  • In the browser, every page registers the same tools with WebMCP (document.modelContext), plus open_token_prediction and open_site_page, so an agent working in the tab can search, read and navigate without scraping the page.

Support