MCP Gateway · Codex
Connect Hugging Face to Codex
Evaluation work depends on picking the right datasets and keeping up with library changes. Through the Speakeasy MCP Gateway, Codex can look up Hub datasets and fetch current Hugging Face documentation while it writes eval harnesses and migrations. Engineers sign in with your identity provider, and Speakeasy can deny any Codex tool call before it runs.
- Category
- AI & search
- Tools
- 10
- Authentication
- OAuth
- Codex
- CLI and ChatGPT desktop app
Behind the MCP Gateway
Run Hugging Face MCP behind the MCP Gateway
One URL for every agent, sign-in through your IdP, RBAC down to the tool, and every Hugging Face call inspected and logged.
Set up
Add Hugging Face to Codex in 4 steps
- 01
Add Hugging Face from the catalog
An admin finds Hugging Face in the MCP Catalog (MCP Gateway > MCP > Catalog) and clicks Add. Speakeasy registers the vendor's server as a remote MCP server and proxies sessions to it. Nothing is copied or re-hosted.
Hugging Face uses OAuth. Servers that support dynamic client registration get their OAuth configuration applied automatically.
Where MCP-scoped policies are enabled, the install offers a recommended risk policy for the new server that detects secrets and PII and inspects tool requests and responses.
- 02
Choose who can use Hugging Face
Give teams the
mcp:connectscope on the Hugging Face server through team access or a role. The organization gateway works out each person's servers from their live access, so granting access adds Hugging Face to their gateway on the next call, and removing it takes it away. - 03
Connect Codex
Add the gateway as an MCP server, then log in.
Run the first command with your organization gateway URL.
Run the second command to sign in.
codex mcp add 'speakeasy' \ --url 'https://<organization host>/mcp/gateway/<organization slug>' codex mcp login 'speakeasy' - 04
Sign in and use Hugging Face from Codex
The first time Codex connects, it opens a browser so the person can sign in to Speakeasy through your identity provider.
Hugging Face needs its own OAuth grant, so each person connects it once. Until they do, Hugging Face calls fail while the rest of the gateway keeps working.
From then on, Hugging Face's tools appear in Codex, limited to what the person's access allows, and every call lands in the gateway's tool logs.
To roll it out to a team, a ChatGPT workspace admin imports the plugin marketplace under Admin > Plugins and sets an installation policy per role. Imported plugins reach nobody until a policy is set.
Use cases
What Codex users do with Hugging Face
Pick datasets for an eval
Codex searches the Hub for datasets that match your task, reads the details of each candidate, and wires the chosen one into the evaluation script.
Migrate to a newer library API
When a Transformers or Gradio upgrade breaks the build, Codex fetches the documentation for the changed API, reading long pages in chunks, and updates each call site.
Check a model's repository details
Before a pinned model goes into production, Codex pulls its repository details so the pull request description records exactly what is being deployed.
Tools
10 tools
Each Hugging Face tool can be allowed or denied per team and role at the MCP Gateway before Codex sees it.
hf_whoami
Hugging Face tools are being used by the authenticated user
space_search
Find Hugging Face Spaces using semantic search. IMPORTANT Only MCP Servers can be used with the dynamic_space toolInclude links to the Space when presenting the results.
hub_repo_search
Search Hugging Face repositories with a shared query interface. You can target models, datasets, spaces, or aggregate across multiple repo types in one call. Use space_search for semantic-first discovery of Spaces. Include links to repositories in your response.
paper_search
Find Machine Learning research papers on the Hugging Face hub. Include 'Link to paper' When presenting the results. Consider whether tabulating results matches user intent.
hub_repo_details
Get details for one or more Hugging Face repos (model, dataset, or space). Auto-detects type unless specified.
hf_doc_search
Search and Discover Hugging Face Product and Library documentation. Send an empty query to discover structure and navigation instructions. Knowledge up-to-date as at 8 April 2026. Combine with the Product filter to focus results.
Control
What Speakeasy can see and stop in Codex
The gateway covers every client the same way. The rest depends on what Codex exposes to Speakeasy.
Tool logs
Gateway
Every Hugging Face call is logged with the person and the client that made it.
Access by role
Gateway
People reach only the Hugging Face tools their team or role allows.
Request and response blocking
Gateway
A risk policy scoped to Hugging Face can reject a tool request before it reaches the server, or withhold a result from Codex. MCP-scoped policies are enabled per organization.
Prompt and tool-call blocking
Codex
Agent hooks send each prompt and tool call to Speakeasy, which allows or denies it before Codex runs it.
User attribution
Codex
An [otel] block in ~/.codex/config.toml sends telemetry to Speakeasy. Export is off by default.
Budgets
Codex
Spend limits can block Codex traffic once a budget is reached.
Device agent and MDM
Codex
The device agent keeps Codex's configuration in sync on managed laptops.
- Codex cloud tasks and GitHub code review run off the device. Neither the plugin nor OpenTelemetry covers them; OpenAI's Compliance Logs record them after the fact.
- The workspace rollout needs a ChatGPT Business, Enterprise or Edu admin.
Questions
Frequently asked questions
Does Speakeasy see Hugging Face calls from Codex cloud tasks?
Can Codex write to Hugging Face repositories?
How do Codex users get Hugging Face?
Are Hugging Face calls from Codex logged?
Connect more