# Zalmir

> Open Beta of a marketplace for AI model inference. One OpenAI-compatible API; listed endpoints are probed on a schedule and each listing shows its last probe time; responses on admitted routes must carry a valid seller signature before the gateway delivers them.

Zalmir routes OpenAI-compatible chat requests to seller endpoints. Probes score each endpoint against captured reference fingerprints (top-k logprob sequences), semantic canaries check output quality, and consecutive-strike flagging reduces transient false positives. On admitted routes the gateway checks the seller's Ed25519 response signature (Zalmir Serving Protocol, ZSP) before delivery and fails the request if the check does not pass. A signature proves the response bytes and the seller's key, not which model or hardware produced them. Seller endpoints can read prompts; confidential (seller-blind) serving is not offered.

## Pages

- [Home](https://zalmir.ai/): marketplace overview with dated verification evidence
- [Models](https://zalmir.ai/models): configured endpoints with dated fidelity scores, verdicts, provider chips, search, and routing and evidence-state filters
- [Model comparison](https://zalmir.ai/models/compare): side-by-side verification scores and verdicts for up to 3 models
- [Verification](https://zalmir.ai/verification): how verification works — probe methodology, scoring, flagging, and what a probe or signature does not prove
- [Status](https://zalmir.ai/status): system health — verification pipeline, gateway, ledger, dashboard; active alerts
- [Clans](https://zalmir.ai/clans): planned seller clans — groups pooling machines into shared listings; not part of the open Beta
- [Trust](https://zalmir.ai/trust): trust boundaries — ZSP signed-response checks, key handling, ledger integrity, and their explicit limits
- [Sell on Zalmir](https://zalmir.ai/sell): model-fit estimates from the hardware you enter; the website cannot detect hardware or start an endpoint, and seller listing is not open in this preview
- [Pricing](https://zalmir.ai/pricing): per-token prices for each model; provider models at the provider's own price with no markup; a top-up fee of 2.5% falling to 1% with lifetime top-ups
- [Docs](https://zalmir.ai/docs): quickstart (account, credits or your own provider key, API key, first request with curl, Python and JavaScript, streaming)
- [API reference](https://zalmir.ai/docs/api): every endpoint with parameters, an example, the response shape and errors
- Guides: [requests and outputs](https://zalmir.ai/docs/guides/requests), [routing and fallbacks](https://zalmir.ai/docs/guides/routing), [privacy controls](https://zalmir.ai/docs/guides/privacy), [your own provider keys](https://zalmir.ai/docs/guides/byok), [credits, pricing and limits](https://zalmir.ai/docs/guides/credits), [errors and headers](https://zalmir.ai/docs/guides/errors), [FAQ](https://zalmir.ai/docs/guides/faq)
- [Console](https://zalmir.ai/dashboard): buyer API keys, credit balance, playground, usage, request receipts
- [Changelog](https://zalmir.ai/changelog): dated development notes and verification research notes
- [Contact](https://zalmir.ai/contact): support and listing reports
- [Terms of Service](https://zalmir.ai/terms): the agreement for the website, Console and API, including credits and the top-up fee
- [Privacy Policy](https://zalmir.ai/privacy): what Zalmir collects, who receives it, how long it is kept and your rights

## API

- Base URL: https://<gateway host supplied by Zalmir>/v1 (OpenAI-compatible). Open Beta: anyone can create an account at https://zalmir.ai/account; the gateway hostname is shown there. Availability depends on deployment.
- Endpoints: POST /v1/chat/completions, POST /v1/responses (stateless), POST /v1/completions (legacy prompt), GET /v1/models, GET /v1/models/{id}, GET /v1/models/{author}/{slug}/endpoints, GET /v1/providers, GET /v1/key, GET /v1/credits, GET /v1/generation?id=, POST /v1/byok/check (checks your saved provider key; nothing charged), GET /openapi.json (the exact runtime contract). All need the bearer key.
- Auth: Bearer buyer key (zmk_…)
- Streaming: `"stream": true` returns OpenAI-compatible SSE (`chat.completion.chunk` events, then `data: [DONE]`). Streaming is buffered: the gateway verifies the whole seller response first and then sends the complete answer, so the first chunk arrives only after the whole answer is ready.
- Successful responses include X-Zalmir-Verified: true, with X-Zalmir-Verification-Scope: response-signature — the gateway checked the seller's ZSP signature, not model identity. Other headers state X-Zalmir-Prompt-Privacy: not-confidential and X-Zalmir-Model-Proof: not-attested.
- Models API: https://zalmir.ai/api/models — public JSON list of configured models with verification metadata; supports `?sort=score|newest|name`, `?provider=`, `?verdict=`, `?q=`, `?limit=&offset=` pagination. Default sort: verification score.

## Verification

- Method: sequence fingerprints (top-k logprobs from a private prompt pool), semantic canaries, baseline normalization, consecutive-strike flagging, multi-anchor cross-checking
- Update cadence: endpoints are probed on a schedule; see each listing's last probe time
- Public data: per-endpoint fidelity score, verdict (faithful / drifted / under review), last probe time

NEW-MODEL RSS FEED: /feed.xml — newly published probe evidence, as it becomes available.

CODING AGENT GUIDES: /docs/agents — untested example configurations for OpenCode, Cline, and Aider (base URL + key + model). Uses the OpenAI chat-completions request shape; specific clients are not yet tested.

EXAMPLE CLIENT SETUPS: /works-with — untested example configurations for Hermes, OpenCode, Cline, Aider, and a generic base URL + key + model recipe. Uses the OpenAI chat-completions request shape; specific clients are not yet tested.