Documentation

Heimdallm is an inference marketplace: independent providers serve models behind one OpenAI-compatible API. You bring an SDK you already use, point it at a new base URL, and pay per token at the price each provider lists. Serving models is open to anyone who can run them — this documentation is the manual for both sides of the market.

Base URL#

All API traffic goes to a single origin:

https://api-staging.impossiblecarrot.cc/api/v1

The API keeps the OpenAI request and response shapes for chat completions, completions, embeddings and transcriptions, so official OpenAI SDKs and most OpenRouter-style clients work by changing baseURL and the API key. Fields Heimdallm adds (such as usage.cost and provider) are extensions that unaware clients can ignore.

Reading these pages as markdown#

Every page here is served as markdown as well as HTML. Append .md to the path (/docs/quickstart.md), or send Accept: text/markdown to the ordinary URL. /llms.txt indexes every page. /AGENTS.md is the short version of this site for a program, covering the API, what a call costs, and how to get a key.

Agent setup#

Install the CLI skill for customer and provider workflows in Hermes, OpenClaw, Claude Code, and Codex. /SKILL.md is the versioned entrypoint.

Buying inference#

New here? The short path:

1. How Heimdallm works — the marketplace in one page: providers, endpoints, models, families, declared versus measured, and the life of a request.
2. Quickstart — sign in, create a key, make your first request.
3. Choosing a model — reading prices, quantizations, measurements and benchmarks to pick between endpoints.
4. Pricing & billing — per-token prices, usage.cost, prepaid credits, holds and spend limits.

Then, as you get real:

• Models & routing — model ids, families, :free / :nitro / :floor, fallback models and provider preferences.
• Production checklist — retries, timeouts, stream discipline, cost control, monitoring.
• Errors & rate limits — the error envelope and every status you can receive.
• Troubleshooting & FAQ — the symptom-to-fix index.

Managing account, money and privacy:

• Your account — balances and holds, the ledger, keys, workspace routing defaults, activity.
• Organizations, workspaces & keys — the account model for teams.
• Deposits — funding the balance: coins, networks, the wallet signature, and what the ledger shows.
• Data & privacy — where requests go, what providers declare, and how to route on it.
• Switching to Heimdallm — from OpenAI, OpenRouter or any OpenAI-compatible client.
• Heimdallm SDKs — the TypeScript and Python clients of the whole API, from inference to video jobs, files, batches, keys and provider routes.
• Client guides: SDKs & frameworks — OpenAI, Anthropic, Vercel AI, LiteLLM, LangChain and LlamaIndex; gateways — Bifrost, LLMGateway and a LiteLLM proxy; chat apps & coding tools — Open WebUI, LibreChat, Cline, Roo Code, Continue and aider.

Capability guides: streaming · tool calling · structured outputs · images & vision · embeddings · documents & files · the playground.

Selling inference#

Providers keep their infrastructure and set their prices; Heimdallm pulls the catalogue, measures the endpoints, routes the traffic and settles the money.

1. Listing your models — the complete provider contract: apply, publish a model list document, pass health probes, get paid monthly in USDC.
2. The provider dashboard — the surface you operate from: sources, pulls, endpoint status, earnings, statements, members, payout settings.
3. Operating a source — pulls and their outcomes, health and lifecycle, capacity reporting, dialects, and the common findings mapped to fixes.
4. Pricing & demand — how price steers routing, what to charge in which units, and how listings grow.
5. A worked example — a real seven-model provider across six rails, configuration included.

API reference#

Every endpoint with parameters, examples and field tables:

• Chat completions · Completions · Embeddings · Rerank · Transcriptions · Speech · Image generation · Video generation · Mesh generation
• Catalogue & status — models, endpoints, providers, families, benchmarks, series, the status board, public stats.
• Key & activity — inspect the key you call with, balances, one generation's billing record, a day of usage.

What the marketplace shows#

Prices, context windows, modalities and data policies are declared by each provider for each endpoint. Uptime, time to first token, throughput and tool-call success are measured by Heimdallm from real traffic and periodic probes — never self-reported. A value a provider does not declare is shown as absent rather than guessed.

Definitions of every term live in the glossary.

Policies#

Pricing states what a request costs and the fee the marketplace takes, the terms of service bind both sides of the market, the privacy policy lists what is stored and for how long, and support says where to ask.