# Hologram Model Hub

> One base URL for open models. Every file is named by the SHA-256 of its bytes and every object by its BLAKE3; fetch from a source that is up, check it yourself, and trust no host, this one included. No account, no key, no SDK.

## The short way

- `export HF_ENDPOINT=https://gethologram.ai` and every tool built on `huggingface_hub` reads from here: `hf download <owner>/<name>`, `snapshot_download`, `from_pretrained`, vLLM, SGLang. llama.cpp reads `MODEL_ENDPOINT` instead. Same commands, same cache; `main` is the indexed revision.
- `ollama pull gethologram.ai/<owner>/<name>:<quant>` for any GGUF in the index; Ollama verifies the SHA-256 itself.
- `oras pull gethologram.ai/<owner>/<name>:latest` for a whole model as an OCI artifact, every layer a raw file whose digest is its SHA-256.
- `GET /api/models?search=qwen&limit=5` to find a model in a few hundred bytes; `GET /api/models/<owner>/<name>/tree/main` for its files, each with `oid`, the SHA-256 it must have.
- `GET /<owner>/<name>/resolve/main/<path>` answers `302` to a source that is up right now; `X-Hub-Source` says which. Put `/via/ipfs`, `/via/modelscope` or `/via/huggingface` in front to pin one.
- Check a whole download with no tool of ours: `curl -s https://gethologram.ai/<owner>/<name>/resolve/main/SHA256SUMS | sha256sum -c`
- MCP: `https://gethologram.ai/mcp`, Streamable HTTP, no key; tools `search_models`, `get_model`, `resolve_file`.

## The one rule

Hash what arrives. Keep a file only if its SHA-256 equals the value from the index (`oid` in the tree, `sha256` in the model object, the layer digest in a manifest). The expected hash never comes from the source that served the bytes, and this server does not verify on read.

## Documentation

### Start

- [Overview](https://gethologram.ai/docs/index.md): What Hologram Hub is, what it replaces, and the one command that proves it.
- [Quickstart](https://gethologram.ai/docs/quickstart.md): Find a model, download one file, and prove the bytes, in three requests and no key.

### Concepts

- [Verification](https://gethologram.ai/docs/verification.md): The one rule that makes a download from anywhere safe, and the three ways to apply it.
- [Names, revisions and addresses](https://gethologram.ai/docs/addresses.md): What a model id means on the hub, why main does not move, and how an address pins a version forever.
- [Sources and failover](https://gethologram.ai/docs/sources.md): Where the bytes come from, how the hub picks a source, and how you pin one.
- [Apps](https://gethologram.ai/docs/spaces.md): Apps that run entirely in your browser, each in its own sealed frame, with every model byte verified before it is used.

### Connect

- [Hugging Face tools](https://gethologram.ai/docs/huggingface.md): Point the tools you already have at the hub with one environment variable.
- [Ollama](https://gethologram.ai/docs/ollama.md): Pull any GGUF in the index straight into Ollama, verified by Ollama itself.
- [OCI artifacts](https://gethologram.ai/docs/oci.md): Pull a whole model as a CNCF ModelPack artifact with oras, modctl, skopeo or crane, every file verified by its digest.
- [MCP](https://gethologram.ai/docs/mcp.md): Give a coding agent the hub as three tools, over one URL, with no key.
- [Objects](https://gethologram.ai/docs/objects.md): The raw, content-addressed index under every dialect, and how to read, mirror or publish to it.

### Reference

- [API reference](https://gethologram.ai/docs/api.md): Every route the hub answers, grouped by what it is for, generated from the OpenAPI document.
- [Errors](https://gethologram.ai/docs/errors.md): Every status the hub answers, the shape it comes in, and what to do about it.
- [Limits and scope](https://gethologram.ai/docs/limits.md): What the hub does not do, how big things can be, and where the edges are.

## Machine-readable

- [OpenAPI 3.1](https://gethologram.ai/openapi.json): the whole endpoint, 50 paths and 52 operations, every example recorded from the live hub
- [agent.md](https://gethologram.ai/agent.md): the whole hub on one screen, for an agent that just arrived
- [Hub descriptor](https://gethologram.ai/.well-known/model-hub.json): today's catalog address and the object routes
- [Agent card](https://gethologram.ai/.well-known/agent-card.json): the same contract as skills
- [Source health](https://gethologram.ai/api/hub/health): which byte sources are up and the order preferred
- [Capabilities](https://gethologram.ai/api/v1/capabilities): what the object server can do and its limits

Index of 2026-09-26: 500 models. Everything here is anonymous and read-only; sign-in exists only for a saved list and model requests.