Hugging Face Inference Endpoints

# Fix list: Hugging Face Inference Endpoints

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/hugging-face-inference-endpoints, the October 2026 research run, assessed 8 October 2026. Grade B, 64.5 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Hugging Face Inference Endpoints: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Payments & pricing, 20 out of 100, up to 10 more on the total

Why it scored 20: No machine payment protocol (0). Hourly prices for every instance published without a login and returned by the unauthenticated `/v2/provider` route (20). No free tier or trial. The docs require a payment method and credits (0). Signup, billing and token creation are browser steps (0).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 2. Reliability, 63 out of 100, up to 7.4 more on the total

Why it scored 63: Better Stack status page at status.huggingface.co with separate Inference Endpoints UI and API components and 90 days of history (20). The API component shows 100 per cent uptime and three scheduled maintenance windows (22 and 30 July, 7 October 2026). The UI was down for 1 hour 37 minutes on 16 July 2026 during a Hub outage, which we count as minor for the API surface (20). No rate limits were found for the management API. The Hub publishes limits per 5-minute window (1,000 API requests for a free user), without saying they cover api.endpoints.huggingface.cloud (5). The docs explain the 503 returned while a replica starts and the `X-Scale-Up-Timeout` header that holds the request, and advise 2 replicas for availability. No backoff guidance or idempotency keys for writes (8). No SLA found in the docs or on the pricing page (0). The service is generally available (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 3. Agent ergonomics, 62 out of 100, up to 6.2 more on the total

Why it scored 62: List endpoints takes `limit` (default 20) and `cursor`, logs take `limit`, `tail` and `line_max_length`, and the MCP server has 19 tools (18). Cursor pagination, filters by state, type, task and tags, sorting, and a time window on logs (18). Errors are documented for the catalogue API and in prose for pause, resume and scale to zero. The management document has no error schema (8). No idempotency keys. Endpoints are addressed by name, updates are a PUT, and the MCP delete tool previews before a confirmed second call. MCP annotations weren't read (8). One-call deploy from the catalogue with tuned defaults, a Python client and the `hf endpoints` CLI. A second-language SDK for endpoint management wasn't found (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 4. Schema & documentation, 73 out of 100, up to 4.4 more on the total

Why it scored 73: Public OpenAPI 3.1 documents for the management API (40 paths, 46 operations) and the catalogue API (3 operations) (25). llms.txt with a Markdown twin of every docs page (10). Every operation has a summary tagged READ, WRITE, PUBLIC or PRO, but only 5 of 46 have a description. The MCP tool table states each tool's purpose and says to call `get_recommended_config` before `create_endpoint` (11). 127 typed schemas with required fields and enums for state, type and accelerator. List filters are comma-separated strings (12). Parameters carry examples. The management document lists only 200 responses, plus 501 on three log routes, while the catalogue document lists 400, 401, 404, 409 and 500 (7). Paths are versioned `/v2` and `/v3` and the document is version 2.0.0. Inference Endpoints has no changelog of its own. The Hub changelog and the docs repository history are the dated record (8).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 5. Security & auth, 83 out of 100, up to 3 more on the total

Why it scored 83: OAuth through huggingface.co with `read-endpoints` and `write-endpoints` scopes, PKCE and dynamic client registration for the MCP server, and revocable fine-grained access tokens for the API. No documented option to send a token in a URL (30). Read and write scopes are separate, endpoints are private by default, organisations on Team and Enterprise plans can require approval of fine-grained tokens, and the MCP delete tool needs `confirm: true`. Other write calls have no confirmation step (17). The service returns the output and logs of the owner's own model, not third-party content (10). An audit log route and MCP tool on Pro, Team or Enterprise plans, plus logs and metrics for every endpoint (12). security.txt valid to 2030 with security@huggingface.co, SOC2 Type 2 stated for the Hub and Inference Endpoints, malware and pickle scanning of repositories. No bug bounty was found (14).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 6. Transparency & trust, 68 out of 100, up to 2.8 more on the total

Made of editorial 68, provenance 67.

Why it scored 68: Closed service under clear terms that name Inference Endpoints, with an Apache-2.0 client (20). The security page says payloads and tokens aren't stored and logs are kept 30 days. The privacy policy dates from 28 March 2023 and gives retention as long as necessary, and a data processing agreement requires an Enterprise subscription (20). Instance deprecations are dated to the month in the pricing table and flagged in the provider API, and the TGI maintenance notice is dated. No deprecation policy for the API was found (10). The privacy policy lists 11 subprocessors with countries and the docs and API list deployment regions (18).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Domain age: huggingface.co, no registry record we could read (0 of 15)
- Endpoint on the vendor's domain: api.endpoints.huggingface.cloud is not on huggingface.co (0 of 15)
- Terms of service: read, states 6 of the 7 things a reader expects, and has 1 clause that costs points (7.1 of 10)

## 7. Maintenance & community, 80 out of 100, up to 1.8 more on the total

Why it scored 80: `huggingface_hub` 2.2.0 was published on PyPI on 8 October 2026 and the docs repository changed on 7 October (30). Releases 1.31.0 to 2.2.0 in the 30 days before, about one a week, and 8 docs commits since 10 July 2026 (20). The GitHub API refused our request, so issue response times went unread. A public changelog for the Hub, a forum and a quota contact form exist (8 of 15). Current official Python client and CLI (15). The client supports Python 3.10 and later and ships weekly. Its CI wasn't read (7).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: GitHub stars, issues and CI for huggingface_hub, because api.github.com answered 403
- unchecked: the Supplemental Terms PDF beyond its first page, which is all our reader extracted
- unchecked: the MCP tool definitions and annotations, because the server needs an OAuth login
- unchecked: the permissions a fine-grained token can carry for Inference Endpoints, which are shown only in the logged-in token settings
- unchecked: security incidents and advisories in the last 12 months, with web search unavailable
- unchecked: the registration date of huggingface.co, because RDAP returned 404
- No SLA, rate limit or idempotency documentation was found for the management API. An Enterprise contract may carry an SLA that isn't published.
- The docs price table and the live `/v2/provider` list disagree on Inferentia2 x1 ($0.75 against $1.95), AWS H200 availability and the RTX PRO 6000 Blackwell, and on the default scale-to-zero period (1 hour against 15 minutes). No deduction was taken.
- The lead was right on the product and interfaces. It omitted the MCP server and the catalogue API. Scale to zero and prices are now read.

## Weaknesses

- No free tier. The docs require a payment method and credits, and replicas are billed while initialising as well as running
- No rate limits, SLA or idempotency keys were found for the management API, and its OpenAPI document lists only 200 responses on 43 of 46 operations
- The docs price table and the live provider list disagree. Inferentia2 x1 is $0.75 in the docs and $1.95 in the API, and AWS H200 is listed in the docs and marked deprecated in the API
- A start from zero replicas takes minutes by the docs' own account, and the proxy answers 503 until a replica is ready
- The Hub outage of 16 July 2026 took the Inference Endpoints UI down for 1 hour 37 minutes

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Call `GET https://api.endpoints.huggingface.cloud/v2/provider` first and pick an instance whose `status` is `available`. The docs table lists types the API marks deprecated or not available
- Send `X-Scale-Up-Timeout: 600` on requests to an endpoint that scales to zero, or handle 503 while the first replica starts
- Set `scaleToZeroTimeout` yourself. The docs give a default of 1 hour and the OpenAPI document says 15 minutes
- Pause or delete an endpoint when the job is done. Billing covers every minute a replica is initialising or running
- Give the agent a fine-grained token or the `read-endpoints` scope unless it must deploy. Endpoints are private by default and take the same Hugging Face token as a bearer

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.