Atlas connects to Hugging Face through the hosted MCP server Hugging Face runs, which gives Atlas model, dataset, Space, and paper search plus semantic documentation search, and any MCP-compatible Gradio Space you enable becomes another tool. Add it with atlas mcp add huggingface --url https://huggingface.co/mcp --header 'Authorization=Bearer $HF_TOKEN', and generate the exact config, including which tools are switched on, at https://huggingface.co/settings/mcp. For inference against Hub models, point Atlas's provider baseURL at https://router.huggingface.co/v1 with the same HF_TOKEN.
What the Atlas and Hugging Face integration gives you
Hugging Face runs a hosted MCP server that gives Atlas model, dataset, Space, and paper search plus semantic documentation search. In 2026 that means Atlas answers questions about the Hub from the Hub, rather than from whatever its training data happened to remember about model names and library APIs.
The tool list is concrete: Model, Dataset, Spaces, and Papers search, plus Documentation Semantic Search and Hub repository details. Semantic documentation search is the one that changes daily work most, because library APIs move faster than any model's training cutoff, and an agent guessing at a function signature wastes a full edit cycle. With the Hugging Face MCP server attached, Atlas looks the answer up. Hugging Face stays the Hub; Atlas stays the terminal agent that queries it.
Adding the Hugging Face MCP server to Atlas
Add Hugging Face's hosted MCP server with 1 command: atlas mcp add huggingface --url https://huggingface.co/mcp --header 'Authorization=Bearer $HF_TOKEN'. The server is hosted at huggingface.co, so nothing installs on your machine, and the bearer token in the header is the only credential involved.
Pass HF_TOKEN by environment reference rather than pasting the literal value into a shell history or a config file. The token determines what Atlas can see on the Hub, including any private repositories your account can reach, so scope it accordingly. Because the endpoint is hosted, there is no container to pull and no Node version to satisfy, which makes the Hugging Face server one of the shortest MCP setups you will do in Atlas.
Generating the exact config at huggingface.co/settings/mcp
Generate the exact config, including which tools are switched on, at https://huggingface.co/settings/mcp. In 2026 that settings page is authoritative about your tool surface, so a tool Atlas cannot see is usually a tool you have not enabled there rather than a fault in the Atlas registration.
Treat the settings page as the control panel for this integration. It shows which of the Hub tools are active for your account and produces the configuration to match. Enabling everything is not automatically the right call: a narrower tool list keeps Atlas's tool listing focused and reduces the chance that the agent reaches for paper search when you wanted repository details. When something is missing, check https://huggingface.co/settings/mcp before you debug anything else.
Turning a Gradio Space into an Atlas tool
Any MCP-compatible Gradio Space you enable shows up as an additional tool on the same Hugging Face server, so Atlas still has just 1 endpoint registered. A Space someone published on the Hub becomes a callable tool in your session without you writing or hosting an MCP server yourself.
The mechanics are simple. Enable the Space from your Hugging Face MCP settings, and the tool appears on the same https://huggingface.co/mcp endpoint Atlas is already connected to, with no second server to register. This is how the Hugging Face integration grows past search: image processing, transcription, evaluation, or any other Gradio Space that speaks MCP becomes something Atlas can call in the middle of a coding task, alongside its file and shell tools.
Running inference against Hub models through the router
To run inference against Hub models, point Atlas's provider baseURL at https://router.huggingface.co/v1 with the same HF_TOKEN. That is a separate concern from the MCP server: the router supplies tokens, while https://huggingface.co/mcp supplies search and documentation tools to the agent.
The two halves of the Hugging Face integration are worth keeping straight. The MCP server at https://huggingface.co/mcp gives Atlas tools. The router at https://router.huggingface.co/v1 gives Atlas a model to think with. You can use either alone. Together, Atlas both searches the Hub and runs on a model hosted through it, and conveniently the same HF_TOKEN authenticates both, so there is only one credential to manage.
Self-hosting the Hugging Face MCP server
The source lives at huggingface/hf-mcp-server if you want to self-host, and as of 2026 it is not published to npm. That detail matters, because an npx one-liner will not work here, so plan to clone and build the repository rather than expecting the usual package-manager install.
Self-hosting is the right call when your organization will not let an agent's tool traffic leave your network, or when you want to pin a specific build. Since huggingface/hf-mcp-server is not published to npm, the setup path is the GitHub repository, and you take on the operational cost of running it. For nearly everyone else the hosted endpoint at https://huggingface.co/mcp with a bearer HF_TOKEN is the shorter and better-maintained option.
Setup
- 01Add Hugging Face's hosted MCP server with atlas mcp add huggingface --url https://huggingface.co/mcp --header 'Authorization=Bearer $HF_TOKEN'
- 02Generate the exact config, including which tools are switched on, at https://huggingface.co/settings/mcp
- 03Confirm the tool surface: the server gives Atlas Model, Dataset, Spaces, and Papers search plus Documentation Semantic Search and Hub repository details.
- 04Enable any MCP-compatible Gradio Space you want, and it shows up as an additional tool on the same server.
- 05To run inference against Hub models, point Atlas's provider baseURL at https://router.huggingface.co/v1 with the same HF_TOKEN.
- 06To self-host instead, build from huggingface/hf-mcp-server, since it is not published to npm.
Frequently asked questions
- how do I connect Atlas to Hugging Face
- Add Hugging Face's hosted MCP server with atlas mcp add huggingface --url https://huggingface.co/mcp --header 'Authorization=Bearer $HF_TOKEN', then confirm your enabled tools at https://huggingface.co/settings/mcp.
- what tools does the Hugging Face MCP server provide
- The server gives Atlas Model, Dataset, Spaces, and Papers search plus Documentation Semantic Search and Hub repository details. Enabled Gradio Spaces appear as additional tools.
- can a Gradio Space be used as an MCP tool
- Yes. Any MCP-compatible Gradio Space you enable shows up as an additional tool on the same Hugging Face MCP server, with no second server to register in Atlas.
- how do I run inference on Hugging Face models with Atlas
- Point Atlas's provider baseURL at https://router.huggingface.co/v1 with the same HF_TOKEN you use for the MCP server. That is inference, separate from the MCP tool surface.
- can I self-host the Hugging Face MCP server
- Yes. The source lives at huggingface/hf-mcp-server, but it is not published to npm, so you build from the repository rather than installing it with npx.
- where do I configure which Hugging Face MCP tools are enabled
- At https://huggingface.co/settings/mcp. The page generates the exact config, including which tools are switched on, which is the first place to check when a tool is missing.
- how does an AI agent search Hugging Face documentation
- Through Documentation Semantic Search on the hosted MCP server. Atlas queries current Hub documentation rather than relying on what a model memorized before its training cutoff.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas with Gemini 2.5 Pro: The Cheap 1M Context Default in 2026
Gemini 2.5 Pro in Atlas: a stable GA id with a 1,048,576 token window at $1.25 per Mtok input, 37 percent cheaper to read with than Gemini 3 Pro.
Atlas with Qwen3 235B-A22B (local via Ollama): Self-Hosted Flagship in 2026
Qwen3 235B-A22B (local via Ollama) in Atlas: roughly 140 GB at 4-bit, Free (self-hosted), 22B active of 235B total, and a brutal memory floor to plan for.
Atlas with Claude Opus 4.1: The Pre Price Cut Opus in 2026
Claude Opus 4.1 runs Atlas on a 200K context at $15 per Mtok input, $75 per Mtok output, with a 32K output ceiling. Setup, cost warnings, and better alternatives.
Atlas for OCaml: A Terminal-Native AI Coding Agent for dune and opam Projects in 2026
Atlas is a terminal-native AI coding agent for OCaml in 2026. It reads dune stanzas and .mli signatures, runs dune runtest behind a prompt, and finishes with ocamlformat.
Atlas with NVIDIA Nemotron 3 Nano 30B A3B in 2026
Nemotron 3 Nano 30B A3B in Atlas, 2026: 3B active parameters at $0.05/$0.20 per Mtok on DeepInfra, free on NVIDIA NIM, up to 1,048,576 tokens on Ollama Cloud.
Atlas with Gemini 3.1 Pro: Setup, Cost, and Tradeoffs in 2026
Run Atlas on Gemini 3.1 Pro in 2026: a 1,048,576 token window at $2 / $12 per Mtok, with real setup steps, the 65,536 output ceiling, and when to switch models.
Atlas with Gemma 2 27B (Ollama): the Free Local Diff Reviewer in 2026
Gemma 2 27B (Ollama) is Google's 2024 flagship open model, 16GB and Free (self-hosted), with 8K tokens (8,192) of context. Use it to review Atlas diffs, not write them.
Atlas with Gemma 3 27B (local via Ollama): Setup, Cost, and Tradeoffs in 2026
Run Atlas on Gemma 3 27B (local via Ollama) in 2026: a 131,072 token context, Free (self-hosted) pricing, single-GPU inference, and the honest tradeoffs.