MLX LM

# Fix list: MLX LM

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/mlx-lm, the October 2026 research run, assessed 8 October 2026. Grade D, 52.2 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 MLX LM: 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. Security & auth, 32 out of 100, up to 11.9 more on the total

Why it scored 32: Read with the tool checklist. The server has no credential of any kind and no flag to add one. It binds 127.0.0.1 by default (5 of 30). `--allowed-origins` defaults to `*`, so a web page can call a server on localhost, and any request can name a `model` or `adapters` path for the server to download or load. `SERVER.md` says local paths must be relative to the start directory, and we found no such check in the code. `--trust-remote-code` is off by default (4 of 20). The server returns model output and parses tool calls without running them. No guidance on untrusted input (8 of 15). Logs at a chosen level, with request bodies at DEBUG, and no per-caller record (5 of 15). A security policy on the Security tab takes reports through GitHub private vulnerability reporting and says findings are published as advisories, the docs and a start-up warning say the server has only basic security checks, and no advisory is published. No bounty is named in the policy (10 of 20).

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.

## 2. Schema & documentation, 37 out of 100, up to 10.2 more on the total

Why it scored 37: Read for the HTTP server, the surface an agent calls. No OpenAPI file or other machine-readable contract. `SERVER.md` says the API is intended to be similar to OpenAI's chat API (5 of 25). No llms.txt. The docs are Markdown files in the repository, 175 lines for the server (5 of 10). Request and response fields each have a one-line purpose and the page opens with the production warning, but `tools`, `seed`, `max_completion_tokens`, `chat_template_kwargs`, the `/v1/completions` route and `/health` are in the code and not in the docs (9 of 20). Types and defaults are given in prose, and the server checks types and ranges in code. We found no `response_format` or JSON schema output in `server.py` (6 of 15). Two curl examples and no documented error responses (5 of 15). Version tags and GitHub release notes, with no changelog file and no version on the API itself beyond the `/v1` path (7 of 15).

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.

## 3. Agent ergonomics, 54 out of 100, up to 7.5 more on the total

Why it scored 54: Read for an API. `max_tokens` defaults to 512, `stop` and `logprobs` shape the output and `usage` reports token counts, but there is no field selection and no structured output mode (15 of 25). `GET /v1/models` lists the MLX models in the Hugging Face cache and can be narrowed to one repository by path. No paging and no token-counting route (10 of 20). Errors are `{"error": "<text>"}` with 400 and a specific message for a bad field, 411 for a missing Content-Length, and 404 for any failure while loading a model or building the prompt. There are no codes and none of it is documented (8 of 20). Generation is stateless and safe to retry, `seed` is accepted, a prompt cache holds ten entries by default and `/health` answers 503 when the generation thread has stopped. No retry guidance (12 of 20). Only `messages` is required and the defaults are conservative. OpenAI clients can call the chat route, and the Python API (`load`, `generate`, `stream_generate`) ships in the same package. No client library in a second language in this repository (9 of 15).

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. Reliability, 66 out of 100, up to 6.8 more on the total

Why it scored 66: Read with the local-software lines, since the package and its server run on the owner's machine. Installs from PyPI (`mlx-lm` 0.32.0) and conda-forge, with Python 3.11 or later stated and macOS and Linux classifiers (20). The Build and Test workflow runs a lint check and the unit tests (21 test files) on a macOS runner for every push to main, and the last eight runs on main passed (25). 151 open issues and 74 open pull requests. The newest bug reports have no reply or one comment, among them two on the server, #1946 (tool calls dropped from the response, 4 October) and #1909 (an uncaught exception that stops the generation thread, 21 September) (15 of 25). Version tags with GitHub release notes, no changelog file, and no breaking changes called out in the ten releases we read. The notes for v0.32.0 did not load (6 of 15). Version 0.32.0, pre-1.0, and the server docs say it is not recommended for production (0).

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.

## 5. Payments & pricing, 60 out of 100, up to 5 more on the total

Why it scored 60: Read with the self-hosted rule. No x402, MPP or L402 in the docs or the source (0). Free under MIT with no account, key or card, and nothing to buy, so 20, 20 and 20 on the last three lines.

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.

## 6. Maintenance & community, 61 out of 100, up to 3.4 more on the total

Why it scored 61: PyPI release 0.32.0 on 1 October 2026 (30). That is the only release in 90 days, the one before being 0.31.3 on 22 April 2026 (0). 127 commits from 82 authors on main since 10 July, with commits on most days of the last week. 151 open issues and 74 open pull requests, and recent bug reports mostly have no reply or one comment. Reply times were not measured (15 of 25). The package is the Python library, current with the release. No other official client (8 of 15). CI on every push, a pre-commit lint job, PyPI publishing by trusted publishing, and dependency floors on recent versions (`mlx>=0.32.2`, `transformers>=5.7.0`). No Dependabot file (8 of 10).

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.

## 7. Transparency & trust, 66 out of 100, up to 3 more on the total

Made of editorial 65, provenance 67.

Why it scored 66: The editorial half. MIT, copyright Apple Inc., all of it public (30). No privacy statement covers the software and nothing states what it sends. In the source the outbound calls are model downloads from Hugging Face (or ModelScope when `MLXLM_USE_MODELSCOPE` is set), uploads only on command, and a calibration text fetched from a GitHub gist by the quantisation tools (12 of 30). Runtime messages say `python -m mlx_lm.<command>` is deprecated, with no date and no written policy (5 of 20). No telemetry, analytics or update check found in the source, so nothing to opt out of, though the README does not say so (18 of 20).

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: apple.com, no registry record we could read (0 of 15)
- Status page: not found (0 of 10)

## 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: the GitHub release notes for v0.32.0, which did not load. The tag (30 September 2026) and the PyPI upload (1 October) were read
- unchecked: reply times on issues and pull requests. Only comment counts on the newest open issues were read
- unchecked: whether Apple's security bounty covers mlx-lm. The repository's policy names GitHub private reporting only
- unchecked: the conda-forge package version. The page answered 200 and was not read
- No terms or privacy document governs the software, so `provenance.terms` and `provenance.privacy` are empty. The pages linked from opensource.apple.com are Apple's website terms and general privacy policy
- Whether the server enforces the documented rule that a local model path is relative to the start directory. We found no check in `server.py` or `utils.py` and did not run the server

## Weaknesses

- `mlx_lm.server` has no API key or other credential option, and `--allowed-origins` defaults to `*`
- A request's `model` and `adapters` fields make the server download or load any Hugging Face repository or local path, with no allow-list (open issue #1892)
- The docs and a start-up warning say the server is not recommended for production because it has only basic security checks
- No OpenAPI file or llms.txt, and `SERVER.md` leaves out `tools`, `seed`, `/health` and the error responses
- One PyPI release in 90 days (0.32.0 on 1 October 2026, the first since 0.31.3 on 22 April), and the version is still 0.x

## 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.

- Keep `mlx_lm.server` on 127.0.0.1 and pass `--allowed-origins` with the origins you trust. There is no API key, and the default answers every origin
- Treat any caller as able to load any model. The `model` and `adapters` request fields accept any Hugging Face repository or local path
- Send `max_tokens` or `max_completion_tokens` when you need more than 512 tokens, the server default
- Read errors as `{"error": "<text>"}` with 400 for a bad field and 404 for a model that failed to load. They are not OpenAI error objects
- Poll `GET /health` before the first request. It answers 503 with `unavailable` when the generation thread has stopped

## 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.