Connector · Data and infrastructure
A teammate in Slack that searches the Hub, on Hugging Face's MCP server.
Chickpea gives a channel a named AI teammate that holds its own Hugging Face connection. You paste a read-scope access token once. The teammate calls Hugging Face's MCP server from then on, so it can search models, datasets, Spaces, and papers without anyone leaving Slack. Ask it which speech models are worth trying, or what this week's papers said, and it answers in the thread.
What people ask it.
The teammate is @oncall, in #eng. Pick a scenario to watch it play.
| Model | Params | Downloads | Licence |
|---|---|---|---|
| solstice-labs/vox-small-en | 240M | 1.2M | Apache 2.0 |
| kettleford/scribe-base | 480M | 610K | Apache 2.0 |
| meridian/asr-mini | 1.1B | 380K | MIT |
| harlowe/listen-v3 | 900M | 210K | CC BY-NC 4.0 |
| duneside/echo-tiny | 75M | 96K | Apache 2.0 |
Find models for a task Read only
@oncall find open weight speech to text models we could self host, with downloads and licence
@oncall which small vision language models have the most downloads this month
Search is a read, so it goes straight through, and a long answer comes back as a table in the thread. Hugging Face says results carry titles, owners, downloads, and links. The licence and the rest of a model card come from the repository details the same server exposes, which is a second read on the same connection.
Find a dataset Read only
@oncall is there a dataset of support tickets with intent labels we could fine tune on
@oncall find datasets of weather time series with at least ten years of history
Datasets are one of the things Hugging Face says the server searches, beside models, Spaces, and papers. The teammate names what it found and what each dataset card says about size, language, and licence, so nobody opens five tabs to rule four of them out.
Read the papers and the documentation Read only
@oncall find papers from the past week on long context evaluation and tell me which three matter
@oncall how do I use LoRA adapters with PEFT
Hugging Face gives both of these as example prompts for its own server, and the second runs its documentation semantic search. Papers are searched the same way as the rest of the Hub, so a date window in the question is a filter on that search rather than a feed the teammate subscribes to.
Run a community Space as a tool Asks first
@oncall run the transcription Space on the three recordings in this thread
@oncall use the image captioning Space on these screenshots
A Space call waits. During a Slack turn Chickpea refuses any MCP tool whose name does not start with an unambiguous read verb unless your message names that action, and a Gradio Space tool never does. Hugging Face says calling a ZeroGPU Space spends the connected account's quota, so the teammate names the Space and asks before it spends anything.
On a schedule Applies at once
@oncall every Monday at 9am Pacific, post the new speech models added to the Hub in the past week
@oncall every Friday at 4pm Eastern, post this week's papers about agent evaluation
A schedule saves on the spot. It needs an explicit time zone, runs as the person who asked and on their Hugging Face connection, and fails with a named reason rather than running if that connection stops resolving.
Connect it
Five steps, from Admin or from Slack.
The credential is asked for once, on a page Chickpea serves, and never in Slack. The link the teammate posts is good for one action, for one teammate, for 24 hours.
- In Chickpea Admin, open the teammate, then Connections, find Hugging Face, and choose Connect. Or ask in Slack,
@oncall connect Hugging Face, and follow the link it posts. - Choose who uses it. Team is one shared token that everyone who can use the teammate acts through. Personal means each member pastes their own, so the Hub sees their account and their quota.
- Make the token. On Hugging Face, open Settings, Access Tokens, choose New token, and pick the read role. Hugging Face says a token with read permissions is what MCP tools need.
- Paste it into the write-only credential field on the page the link opens. It begins
hf_, it goes out as a bearer token, and it is stored once on your deployment. Nothing is pasted into Slack. - Chickpea connects to huggingface.co/mcp before it saves anything and records the tools the server answered with. The teammate is handed that list, which can later be narrowed to fewer tools but never widened. Then ask something, like
@oncall find open weight speech to text models.
A connection belongs to one teammate. Giving a second teammate Hugging Face means a second connection with its own token, so the growth teammate never inherits what the oncall teammate can reach. Which tools the server offers depends on the account behind the token, because Hugging Face's own MCP settings page is where you add Spaces and switch the extra tools on.
There is no built-in Chickpea skill for Hugging Face, and none is needed. The tools arrive from Hugging Face's MCP server. Chickpea connects before it saves anything, records what the server answered with, and exposes only those tools. Hugging Face says most Hub work is done by one built-in tool, hf_fs, which navigates the Hub and runs semantic search over its documentation and Spaces, and that you switch on Contribute Repos, Sandboxes, and Run and Manage Jobs from its settings page. Hugging Face documents the server, its built-in tool, and the Spaces you can add.
Beside Hugging Face
What the same teammate usually holds too.
Each is its own connection, granted to the teammate alone. All 37 connectors.
ExaA model found on the Hub, beside what the web has written about it
FirecrawlThe model card from the Hub, the whole docs site behind it from Firecrawl
SupabaseA model picked on the Hub, the rows it would run against in Supabase
PostHogWhat the new model shipped into, checked against what people did next
Questions people ask
What is the Hugging Face MCP server?
A server Hugging Face runs that speaks the Model Context Protocol, so an AI assistant can search and explore the Hub through it without writing HTTP calls of its own. Hugging Face says it covers models, datasets, Spaces, papers and more, adds natural language search over its documentation, and can run community tools built as MCP-compatible Gradio Spaces. Chickpea connects to that same server, which is why a teammate in a Slack channel can use it with no client installed anywhere.
What is the URL for the Hugging Face MCP server?
Hugging Face documents the server at huggingface.co/mcp, which doubles as the settings page where you sign in and pick a client. In Chickpea you never type it. The Hugging Face preset carries that URL on the streamable HTTP transport, and a connection made from the preset cannot be pointed at another address.
Does it use Hugging Face's MCP server or the API?
The MCP server. Hugging Face is a vendor MCP lane here: the deployment connects to huggingface.co/mcp over streamable HTTP and calls the tools that server answered with. The connector never touches the Hugging Face REST API or an Inference Providers endpoint, and the connection can reach no host but that one.
Which Hugging Face token does it need?
An access token that begins hf_, with the read role. Hugging Face warns that MCP tools need a valid token with read permissions. Make one under Settings, Access Tokens, choose New token, and pick read. Hugging Face describes that role as giving read access to repositories the account could read already, meaning public repositories plus the private ones you or your organizations own. The catalog entry for this preset says the same thing.
Where does the token live?
On your deployment, outside the teammate record, in a write-only field. It never enters the model's context, the tool arguments the model writes, or Slack. The model picks a connection by id and the token goes into the Authorization header at the last step on the way out. Disconnect the account and the stored secret is removed, and any schedule that depended on it stops.
Can it upload a model, change a repo, or delete anything?
No, for two separate reasons. The preset asks for a read token, and Hugging Face defines that role as read access to repositories the account could already read, so it carries no write access to any repository. Separately, during a Slack turn Chickpea refuses any MCP tool whose name does not start with an unambiguous read verb unless your own message names that action. Hugging Face does publish tools for contributing to repositories, running Jobs, and using Sandboxes, switched on from its own MCP settings page, and a read token is not the credential for those.
Is the Hugging Face MCP server free?
Hugging Face publishes no price for the server and no rate limit for it, so this page puts no number on either. What it does put in writing is the quota behind Spaces: calling a ZeroGPU Space as a tool spends the connected account's quota, and running out means waiting or subscribing to PRO, which Hugging Face says carries 40 minutes of daily quota, eight times what a free account gets. Searching the Hub is not a Space call. Chickpea itself is Apache 2.0 and runs on your own Cloudflare account, so there are no seats and no metering on this side.
Can it run a Space as a tool?
If the Space supports it and the account has added it. Hugging Face lets you turn any public Space carrying the MCP badge into a callable tool from its MCP settings, and it has a Dynamic Spaces option that lets a client discover and call MCP-compatible Spaces at runtime. Whatever the account has switched on shows up in the tool list Chickpea records when the connection is made. A Space's output is retrieved content, which the teammate treats as material to read and never as an instruction, so a transcript telling it to email somebody is data it can quote back to you and nothing more.
Is this the Hugging Face Agents Course?
No. That is Hugging Face's free course on building agents, and it is what most searches for a Hugging Face AI agent are after. This page is about a connector. The teammate already exists in your Slack workspace, and connecting Hugging Face gives it the Hub in the channel where somebody asked the question. There is no course and nothing to install.
Can I use the Hugging Face MCP server from Claude Code or Cursor instead?
Yes, and Hugging Face documents that route first. Its settings page generates a ready-to-copy configuration for Codex, Cursor, VS Code extensions, Zed, ChatGPT, and Claude Desktop, among others. That puts the server in one person's editor. A teammate puts it in a channel, with a stored connection, saved instructions, schedules, and every search visible in the thread where it was asked for. Plenty of teams run both.
Try it
Give a channel a Hugging Face teammate.
One deploy into your own Cloudflare account, connect Slack, paste one token, ask. Apache 2.0, no seats, no metering.
Or run it on Node. The setup guide covers both.
Last checked
Facts on this page were checked against the product and Hugging Face's own pages on September 17, 2026. This page is also available as markdown at /connectors/huggingface.md.