> ## Documentation Index
> Fetch the complete documentation index at: https://docs.formal.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Hugging Face

> Search and work with models, datasets, Spaces, and Hugging Face resources.

Knowledge Publisher: Hugging Face

Route Hugging Face MCP traffic through Formal to control and log what AI agents do in Hugging Face:

* **Block** tools you don't want agents to call, such as `dataset_search`.
* **Mask** sensitive data in tool results, such as emails and phone numbers in `model_search` responses.
* **Keep credentials at the Connector**, so users and agents never see Hugging Face tokens.
* **Log** every tool call with the user, the MCP client and its parameters.

Cloud agents

Laptop / VM

Endpoint

[https://huggingface.co/mcp](https://huggingface.co/mcp)

Connector

![](https://docs.formal.ai/static/brands/hugging-face.svg)

## 0. Prepare your Connector

The Connector needs an MCP listener and a hostname with wildcard DNS and TLS.
Set them up once by following [Deploy the Connector](https://docs.formal.ai/docs/guides/core-concepts/connectors/mcp-gateway.md#deploy-the-connector). Every MCP server reuses them.

## 1. Add Hugging Face MCP server to your Resources

> [tab] UI
>
> 1. Open the [Admin Catalog](https://app.formal.ai/catalog) and click **Add MCP**.
> 2. Select the **Hugging Face** template.
> 3. Enter a DNS-label **Name**, such as `hugging-face`.
> 4. Choose the **Space** containing your Connector, such as `default`.
> 5. Configure upstream authentication as described below, then save.
>
> If no Connector is in the selected Space, the **Configuration** tab shows **MCP clients can't connect directly yet**.
> The reason reads: `No Connector is in this server's Space.`
>
> Adding Hugging Face to the Admin Catalog

> [tab] Terraform
>
> Set `space_id` to the ID of the Space containing your Connector.
>
> ```hcl theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> resource "formal_resource" "hugging_face" {
>   name                    = "hugging-face"
>   technology              = "mcp"
>   hostname                = "huggingface.co"
>   space_id                = "<SPACE_ID>"
>   native_users_v3_enabled = true
>   port                    = 443
>   tags = {
>     "mcp-endpoint" = "/mcp"
>   }
> }
>
> resource "formal_resource_oauth_settings" "hugging_face" {
>   resource_id = formal_resource.hugging_face.id
> }
> ```

> [tab] Pulumi
>
> Set `spaceId` to the ID of the Space containing your Connector.
>
> ```ts theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> import * as formal from "@formalco/pulumi";
>
> const resource = new formal.Resource("mcp", {
>   name: "hugging-face",
>   technology: "mcp",
>   hostname: "huggingface.co",
>   port: 443,
>   spaceId: "<SPACE_ID>",
>   tags: { "mcp-endpoint": "/mcp" },
>   nativeUsersV3Enabled: true,
> });
>
> new formal.ResourceOauthSettings("mcpOAuth", {
>   resourceId: resource.id,
> });
> ```

> [tab] API
>
> Create a Resource with the MCP endpoint and its Space ID.
>
> ```bash theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> export FORMAL_API_KEY="<YOUR_API_KEY>"
> RESOURCE_RESPONSE=$(curl --fail --silent --show-error \
>   --request POST \
>   https://api.formal.ai/core.v1.ResourceService/CreateResource \
>   --header "x-api-key: ${FORMAL_API_KEY}" \
>   --header "Content-Type: application/json" \
>   --data-binary @- <<JSON
>   {
>     "name": "hugging-face",
>     "hostname": "huggingface.co",
>     "port": 443,
>     "technology": "mcp",
>     "spaceId": "<SPACE_ID>",
>     "tags": [
>       {
>         "key": "mcp-endpoint",
>         "value": "/mcp"
>       }
>     ]
>   }
> JSON
> )
> RESOURCE_ID=$(printf '%s' "${RESOURCE_RESPONSE}" | jq -r '.resource.id')
> curl --fail --silent --show-error \
>   --request POST \
>   https://api.formal.ai/core.v1.ResourceService/CreateResourceOAuthSettings \
>   --header "x-api-key: ${FORMAL_API_KEY}" \
>   --header "Content-Type: application/json" \
>   --data-binary @- <<JSON
>   {
>     "resourceId": "${RESOURCE_ID}"
>   }
> JSON
> ```

## 2. Setup Authentication

> [tab] Linked Accounts (Recommended)
>
> Each user links their own Hugging Face account once in the Employee Catalog, and the Connector sends that user's token to Hugging Face.
>
> ![Linked Accounts selected under How does Formal authenticate to this resource?](https://mintcdn.com/formal/vzPMpIcsF4BYmdx8/assets/images/mcp-servers/auth-linked-accounts.png?fit=max\&auto=format\&n=vzPMpIcsF4BYmdx8\&q=85\&s=dc448db2a373faa6e1bf49fd5ffaf7c4)
>
> Under **OAuth app**, choose one of the two options.
>
> #### 1. Formal's App (default)
>
> Formal has a registered OAuth app with Hugging Face, so you register nothing.
> Users see Formal on Hugging Face's consent screen.
>
> #### 2. Your OAuth App
>
> Users see your app's name on Hugging Face's consent screen.
>
> > [accordion] Set up your own Hugging Face OAuth app
> >
> > See [Hugging Face OAuth documentation](https://huggingface.co/docs/hub/oauth) and [register an app](https://huggingface.co/settings/applications/new).
> >
> > **1. Create an OAuth app**
> >
> > Open [Connected Apps settings](https://huggingface.co/settings/applications/new) to create a new OAuth app. Enter a name for it.
> >
> > **2. Add the redirect URI**
> >
> > Paste Formal's **Redirect URI** into the app's redirect URIs field. It must match exactly.
> >
> > **3. Choose scopes**
> >
> > Select `openid`, `profile`, and `read-mcp`. Add `read-repos`, `jobs`, `contribute-repos`, or `inference-api` only if your tools need them.
> >
> > **4. Copy the credentials**
> >
> > Save the app, then copy its **Client ID** and **Client secret**. You can skip the client secret. Formal then authenticates with PKCE alone.
> >
> > **5. Enter the app in Formal**
> >
> > > [tab] UI
> > >
> > > Back in the Admin Catalog, enter the **Client ID** and **Client secret**, then save.
> > > Users then link their accounts on the Hugging Face server's page in the [Employee Catalog](https://catalog.formal.ai).
> > >
> > > ![Your OAuth App selected in the Admin Catalog, with the Redirect URI, Client ID and Client secret fields](https://mintcdn.com/formal/R-3CwVgs8IU6liji/assets/images/mcp-servers/your-oauth-app.png?fit=max\&auto=format\&n=R-3CwVgs8IU6liji\&q=85\&s=00584678f4f2696805678de060eaea0a)
> >
> > > [tab] Terraform
> > >
> > > Use these OAuth settings instead of the ones from step 1.
> > >
> > > ```hcl theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> > > variable "hugging_face_oauth_client_secret" {
> > >   type      = string
> > >   sensitive = true
> > > }
> > >
> > > resource "formal_resource_oauth_settings" "hugging_face" {
> > >   resource_id              = formal_resource.hugging_face.id
> > >   upstream_oauth_mode      = "pre_registered"
> > >   client_id                = "<YOUR_CLIENT_ID>"
> > >   client_secret_wo         = var.hugging_face_oauth_client_secret
> > >   client_secret_wo_version = 1
> > > }
> > > ```
> >
> > > [tab] Pulumi
> > >
> > > Use these OAuth settings instead of the ones from step 1.
> > > Store the secret with `pulumi config set --secret huggingFaceOauthClientSecret`.
> > >
> > > ```ts theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> > > import * as pulumi from "@pulumi/pulumi";
> > >
> > > const config = new pulumi.Config();
> > >
> > > new formal.ResourceOauthSettings("mcpOAuth", {
> > >   resourceId: resource.id,
> > >   upstreamOauthMode: "pre_registered",
> > >   clientId: "<YOUR_CLIENT_ID>",
> > >   clientSecret: config.requireSecret("huggingFaceOauthClientSecret"),
> > > });
> > > ```
> >
> > > [tab] API
> > >
> > > Use this call instead of the `CreateResourceOAuthSettings` call from step 1.
> > > The OAuth app is fixed at creation.
> > >
> > > ```bash theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> > > curl --fail --silent --show-error \
> > >   --request POST \
> > >   https://api.formal.ai/core.v1.ResourceService/CreateResourceOAuthSettings \
> > >   --header "x-api-key: ${FORMAL_API_KEY}" \
> > >   --header "Content-Type: application/json" \
> > >   --data-binary @- <<JSON
> > >   {
> > >     "resourceId": "${RESOURCE_ID}",
> > >     "upstreamOauthMode": "pre_registered",
> > >     "clientId": "<YOUR_CLIENT_ID>",
> > >     "clientSecret": "<YOUR_CLIENT_SECRET>"
> > >   }
> > > JSON
> > > ```
> >
> > > [!WARNING]
> > >
> > > Switching between Formal's App and your OAuth app unlinks every account. Users must link again.

> [tab] Shared Credentials
>
> An administrator stores one Hugging Face token in Formal, and the Connector sends it on every user's requests without showing it to them.
>
> ![Shared Credentials selected under How does Formal authenticate to this resource?](https://mintcdn.com/formal/R-3CwVgs8IU6liji/assets/images/mcp-servers/auth-shared-credentials.png?fit=max\&auto=format\&n=R-3CwVgs8IU6liji\&q=85\&s=91032e202518088292ee1cf39451b9ef)
>
> Hugging Face's MCP endpoint accepts an [User Access Token](https://github.com/huggingface/hf-mcp-server) as a Bearer token.
> Create one at [huggingface.co/settings/tokens](https://huggingface.co/settings/tokens).
>
> 1. Open this server from the [Admin Catalog](https://app.formal.ai/catalog).
> 2. Select **Configuration**, then **Authentication**.
> 3. Choose **Shared Credentials**, then select **Add credential**.
> 4. Enter the credential, make it the default, and select **Save authentication**.
>
> See [Native Users](https://docs.formal.ai/docs/guides/core-concepts/resources/native-users.md).
>
> For HTTP Bearer authentication, enter the raw token. The Connector adds the `Bearer ` prefix.
>
> In the Terraform tab, set `native_users_v3_enabled = true` on the Resource and `upstream_oauth_mode = "none"` on its OAuth settings.
>
> ```hcl theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> variable "mcp_shared_credential" {
>   type      = string
>   sensitive = true
> }
>
> resource "formal_native_user_v3" "mcp_shared_credential" {
>   resource_id = formal_resource.hugging_face.id
>   label       = "shared-credential"
>
>   http_bearer {
>     header = "Authorization"
>     token {
>       literal_wo         = var.mcp_shared_credential
>       literal_wo_version = 1
>     }
>   }
> }
>
> resource "formal_resource_native_user_selection" "mcp_shared_credential" {
>   resource_id = formal_resource.hugging_face.id
>   cel         = "\"${formal_native_user_v3.mcp_shared_credential.id}\""
> }
> ```

> [tab] Pass Through
>
> Each user sends their own Hugging Face sign-in or token from their MCP client, and Formal forwards it unchanged without storing it.
>
> ![Pass Through selected under How does Formal authenticate to this resource?](https://mintcdn.com/formal/R-3CwVgs8IU6liji/assets/images/mcp-servers/auth-pass-through.png?fit=max\&auto=format\&n=R-3CwVgs8IU6liji\&q=85\&s=c14ae61de2dba8b567626cba03aae2cf)
>
> Nothing in the request identifies the Formal user. Formal attributes these requests only when they come through the [Formal Endpoint](https://docs.formal.ai/docs/guides/core-concepts/connectors/mcp-gateway.md#identity-headers). Pass-through servers have no Connector address, so cloud agents can't use them.
>
> For OAuth, each MCP client signs in to Hugging Face itself.

## 3. Route Endpoint traffic

Auth is managed at the Connector level when you use Linked Accounts or Shared Credentials, so the Endpoint needs to forward the MCP traffic to it.
Laptops and VMs with the Formal Endpoint keep using `https://huggingface.co/mcp`, and a network rule sends that traffic to the Connector.

> [tab] UI
>
> 1. Go to [Network Rules](https://app.formal.ai/network-rules) and create a rule named `hugging-face-mcp`.
> 2. Paste this expression, set the rule to **Active**, then save:
>
> ```javascript theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> {
>   "condition": {
>     "pre_tcp": process.is_agent,
>     // Every MCP server added to Formal: process.is_agent && resource.technology == "mcp"
>     "pre_tls": process.is_agent && hostname == "huggingface.co",
>     "post_tls": true
>   },
>   "outputs": {
>     "forward_to_connector": true
>   }
> }
> ```

> [tab] Terraform
>
> ```hcl theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> resource "formal_network_rule" "hugging_face_mcp" {
>   name        = "hugging-face-mcp"
>   description = "Forward Hugging Face MCP traffic from agents to the Connector"
>   status      = "active"
>
>   # Every MCP server added to Formal: set pre_tls to process.is_agent && resource.technology == "mcp"
>   cel_expression = <<-EOT
>     {
>       "condition": {
>         "pre_tcp": process.is_agent,
>         "pre_tls": process.is_agent && hostname == "huggingface.co",
>         "post_tls": true
>       },
>       "outputs": {
>         "forward_to_connector": true
>       }
>     }
>   EOT
> }
> ```

> [tab] Pulumi
>
> ```ts theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> import * as formal from "@formalco/pulumi";
>
> new formal.NetworkRule("huggingFaceMcp", {
>   name: "hugging-face-mcp",
>   description: "Forward Hugging Face MCP traffic from agents to the Connector",
>   status: "active",
>   // Every MCP server added to Formal: set pre_tls to process.is_agent && resource.technology == "mcp"
>   celExpression: `
>     {
>       "condition": {
>         "pre_tcp": process.is_agent,
>         "pre_tls": process.is_agent && hostname == "huggingface.co",
>         "post_tls": true
>       },
>       "outputs": {
>         "forward_to_connector": true
>       }
>     }
>   `,
> });
> ```

> [tab] API
>
> ```bash theme={"languages":{"custom":["/languages/cel.json","/languages/rego.json"]}}
> export FORMAL_API_KEY="<YOUR_API_KEY>"
> # Every MCP server added to Formal: set pre_tls to process.is_agent && resource.technology == "mcp"
> CEL_EXPRESSION=$(cat <<'CEL'
> {
>   "condition": {
>     "pre_tcp": process.is_agent,
>     "pre_tls": process.is_agent && hostname == "huggingface.co",
>     "post_tls": true
>   },
>   "outputs": {
>     "forward_to_connector": true
>   }
> }
> CEL
> )
> jq -n --arg cel "${CEL_EXPRESSION}" '{
>   name: "hugging-face-mcp",
>   description: "Forward Hugging Face MCP traffic from agents to the Connector",
>   status: "active",
>   celExpression: $cel
> }' | curl --fail --silent --show-error \
>   --request POST \
>   https://api.formal.ai/core.v1.DesktopService/CreateDesktopRoutingRule \
>   --header "x-api-key: ${FORMAL_API_KEY}" \
>   --header "Content-Type: application/json" \
>   --data-binary @-
> ```

The rule matches [known agents](https://docs.formal.ai/docs/guides/network-rules.md#known-agent-identities) only.
See [Identity headers](https://docs.formal.ai/docs/guides/core-concepts/connectors/mcp-gateway.md#identity-headers).

Cloud agents such as Claude, ChatGPT and Cursor Cloud Agents don't run the Endpoint.
Point them at this server's Connector address, `https://hugging-face.<connector-hostname>/mcp`.
Find it in the [Admin Catalog](https://app.formal.ai/catalog): open Hugging Face, then copy **Connector address** on the **Configuration** tab.

## 4. Create policies

### Tools

The Admin Catalog lists these tools for Hugging Face:

| Tool | Description |
| - | - |
| `model_search` | Search models on the Hugging Face Hub. |
| `dataset_search` | Search datasets. |
| `space_search` | Search Spaces. |
| `paper_search` | Search research papers. |
| `hub_repo_details` | Retrieve repository metadata. |
| `hf_doc_search` | Search Hugging Face documentation. |

## 5. Check the logs

Open [Logs](https://app.formal.ai/logs) and search for `resource.name:hugging-face`.
Each Hugging Face tool call is one `request` log with the user, the MCP client and the tool.
Open a log to see its tool parameters and the policies it triggered.

Hugging Face tool calls in Logs, including a masked response and a blocked write

See [Logs](https://docs.formal.ai/docs/guides/observability/logs.md) for search syntax and filters.

## Troubleshooting

> [accordion] The agent says Formal has no current sign-in
>
> The agent relays: `Formal has no current hugging-face sign-in for you. Link or re-link your hugging-face account in Formal…`
>
> **Cause:** The user never linked Hugging Face, or the token expired and couldn't be refreshed.
>
> **Fix:** Open the server in the [Employee Catalog](https://catalog.formal.ai) and link Hugging Face again.

> [accordion] The agent says Hugging Face rejected your credentials
>
> The agent relays: `hugging-face rejected your credentials. Link or re-link your hugging-face account in Formal…`
>
> **Cause:** Hugging Face no longer accepts the linked token, for example after the user revoked access in Hugging Face.
>
> **Fix:** Link Hugging Face again from the [Employee Catalog](https://catalog.formal.ai).
>
> With Shared Credentials, the message asks the user to contact an administrator. Replace the token in the server's **Shared Credentials**.

> [accordion] Laptop traffic doesn't appear in Logs
>
> **Cause:** The Endpoint isn't running, or the network rule doesn't match the client.
>
> **Fix:** Check the Endpoint is connected. The rule above only matches [known agents](https://docs.formal.ai/docs/guides/network-rules.md#known-agent-identities). Other clients, such as a browser, reach Hugging Face directly unless the rule also matches them. See [Network rules](https://docs.formal.ai/docs/guides/network-rules.md).

> [accordion] MCP clients can't connect directly yet
>
> **Cause:** The **Configuration** tab lists the reason, most often `No Connector is in this server's Space.`
>
> **Fix:** Edit the server and choose the Space containing its Connector.

## Next Steps

> [card] [MCP Gateway](https://docs.formal.ai/docs/guides/core-concepts/connectors/mcp-gateway.md)
>
> Configure Connector routing and downstream access

> [card] [All MCP servers](https://docs.formal.ai/docs/guides/core-concepts/connectors/mcp-servers/index.md)
>
> Browse every MCP server template


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.