# Fix list: llama.cpp From Anchor Terminal's listing at https://www.anchorterminal.com/tools/llama-cpp, the October 2026 research run, assessed 3 October 2026. Grade C, 60.2 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 llama.cpp: 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. Schema & documentation, 47 out of 100, up to 8.6 more on the total Why it scored 47: No spec of its own. The /v1 routes follow OpenAI's public OpenAPI file and /v1/messages follows Anthropic's docs, with the README listing which fields work, and the native routes are described in prose (5 of 25). No llms.txt (llama.app/llms.txt returns 404). The server reference is one Markdown file of 2,301 lines in the repository, readable raw (5 of 10). Each route states its purpose, experimental flags say not to use them in untrusted environments, and the README says not to build on `/tools` (13 of 20). Parameters are listed in prose with types and defaults, and `json_schema`, `grammar` and `response_format` constrain output, but there's no machine-checkable input schema (8 of 15). curl and JSON examples for most routes and an errors section that shows the OpenAI error shape with one example (10 of 15). Semver tags since August 2026 and nightly builds with generated notes, but no changelog file, and the REST changelog hasn't changed since b4599 (6 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. ## 2. Security & auth, 52 out of 100, up to 8.4 more on the total Why it scored 52: Read with the tool checklist, credential model first. Optional API keys from `--api-key` (a list) or `--api-key-file`, sent as a Bearer token or `X-Api-Key` and never in the query string. They have no scopes, change only with a restart and are off by default. TLS is built in, and the server binds 127.0.0.1 by default (15 of 30). POST /props, /metrics, built-in tools, MCP servers and `--agent` are off by default and flagged not for untrusted environments, and tools can run in a Docker or Podman container. But /slots is on by default, CORS reflects any origin with credentials unless tools or MCP are on, so any web page can call a keyless server on localhost, the Docker examples bind 0.0.0.0 with no key, and every key can do everything (8 of 20). SECURITY.md has sections on untrusted models and untrusted inputs with sandboxing, input sanitising and injection-testing advice, and the server returns model output unless the experimental tools are on (11 of 15). Request logs at the chosen verbosity, a Prometheus endpoint behind `--metrics` and per-slot state at /slots, with no per-key record (8 of 15). SECURITY.md with a scope and a 90-day window, ten published GitHub advisories with CVEs and fixed builds (four from January to March 2026), and a periodic AI code scan whose prompts are public. Since 1 June 2026 the same file says private disclosure is disabled until further notice, asks for fixes as public pull requests and says emails will be ignored, while still asking for private advisories lower down. No security.txt or bug bounty (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. ## 3. Reliability, 64 out of 100, up to 7.2 more on the total Why it scored 64: Read with the local-software lines, since llama-server runs on the owner's machine with no hosted service. Official installs through the llama.app script, winget, Homebrew, MacPorts, Nix, conda-forge and Docker images on ghcr.io, plus binaries for every nightly build, with each backend's requirements in docs/build.md (20). 37 workflows run on pushes to master across CPU, CUDA, Metal, Vulkan, SYCL, sanitiser and server builds, and the last five finished runs of the server sanitiser workflow on master passed, with three queued. We didn't check the rest (22 of 25). 868 open issues, held down by a stale bot that closes inactive issues without a bug, security or roadmap label after 44 days. New reports are labelled bug-unconfirmed, and this week's crash reports (#29811, #29783, #29780, #29774) range from 14 comments to none (15 of 25). Semver since v0.1.0 on 17 August 2026, with a written rule that a breaking change to llama.h bumps the major version, but releases are bare tags with no notes, nightly builds carry generated commit lists, and the server's REST changelog (#9291) stops at b4599 (7 of 15). Version 0.5.0, pre-1.0 (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. ## 4. 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. ## 5. Agent ergonomics, 73 out of 100, up to 4.4 more on the total Why it scored 73: Read for an API. `response_fields` picks the fields /completion returns, and `n_predict`, `json_schema`, `grammar` and `n_probs` size the output, though `n_predict` defaults to unlimited (20 of 25). Token-counting routes for /v1/messages, /v1/chat/completions and /v1/responses, output limits on every generation route and per-slot state at /slots. Model lists aren't paged (16 of 20). Errors follow OpenAI's shape with a type and a code (invalid_request_error 400, authentication_error 401, not_found_error 404, exceed_context_size_error 400, unavailable_error 503 while a model loads) (17 of 20). Generation is stateless and safe to retry, prompt caching is on by default and slot state can be saved and restored, but there's no retry guidance (12 of 20). A model and a port make a working server, routes need few fields, and OpenAI and Anthropic clients work against it, but there's no official client library in any language (8 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. ## 6. Transparency & trust, 60 out of 100, up to 3.5 more on the total Made of editorial 66, provenance 53. Why it scored 60: The editorial half. MIT, all of it public (30). We found no privacy policy on llama.app or in the repository and no statement of what the software sends. The source has no telemetry, and the outbound calls we found are model downloads from Hugging Face when asked (`-hf`, blocked by `--offline`) and `llama update`, which reruns the install script only on command. SECURITY.md advises sandboxing for private data (12 of 30). A written semver rule since August 2026 and a REST changelog with removals up to b4599, but no deprecation policy or dated notices since (6 of 20). No telemetry, analytics or update check in the source, so nothing to opt out of, though nothing on llama.app or in the README says 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: llama.app, no registry record we could read (0 of 15) - Status page: not found (0 of 10) - security.txt: not found (0 of 10) ## 7. Maintenance & community, 81 out of 100, up to 1.7 more on the total Why it scored 81: Nightly build b11375 on 3 October 2026 and release v0.5.0 on 23 September (30). 1,005 tagged nightly builds since b9873 on 5 July and eight semver releases since 17 August (20). 868 open issues and 1.7k open pull requests, 1,503 commits on master in 90 days from 391 authors, a bot that links related issues on each new one, and a stale bot. Some recent crash reports have long threads (#29811, 14 comments) and others no reply yet. GitHub's issue search is closed to our reader, so reply times are unchecked (15 of 25). No official client library for the server. The C API in llama.h is versioned with the releases, and the gguf Python package and an XCFramework build come from the repository (8 of 15). CI on every push to master across backends and a check on the vendored code, with no Dependabot (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. ## Deductions Each comes off the total. A fixed and documented problem counts for less at the next check. - 2026-03-26. GHSA-j8rj-fmpv-wcxw (CVE-2026-34159, 9.8 at NVD), unauthenticated code execution through a GRAPH_COMPUTE bypass in the RPC backend, the most serious of four advisories published between January and March 2026 (the others a llama-server out-of-bounds write through a negative `n_discard` and two GGUF integer overflows). All were fixed in named builds and published as advisories, SECURITY.md says not to expose the RPC server or llama-server to untrusted networks, and the newest is more than six months old, -1. https://github.com/ggml-org/llama.cpp/security/advisories/GHSA-j8rj-fmpv-wcxw; https://github.com/ggml-org/llama.cpp/security ## 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: reply times on issues, since GitHub's issue search is closed to our reader - unchecked: the state of CI workflows on master other than the server sanitiser workflow - unchecked: the legal entity behind llama.app and ggml.ai after the Hugging Face acquisition, and its date, since neither site names one - Whether the private disclosure programme will return, and how reports filed under it before 1 June 2026 were handled - unchecked: first release date. The earliest b-tag we dated, b1046, is from 24 August 2023, and the master-* tags before it weren't dated ## Weaknesses - API keys are off by default and CORS reflects any origin with credentials, so a web page can call a keyless server on localhost - No OpenAPI file of its own, and the REST API changelog stops at b4599 - Private security disclosure disabled since 1 June 2026, with fixes asked for as public pull requests - Pre-1.0 (0.5.0), and semver releases are bare tags with no notes - No official client library, and `n_predict` defaults to unlimited ## 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. - Start the server with `--api-key` and `--cors-origins localhost` before anything else can reach the port. Both are off by default - Pass `n_predict` or `max_tokens`. Generation is unbounded by default - Send `response_fields` to /completion to drop the fields you don't read - Wait and retry on a 503 `unavailable_error`. The model is still loading - Read the server README of the build you run. Behaviour changes between nightly builds without a changelog entry ## What the review panel asked for - notes on semver releases - a current REST changelog - restore private disclosure - narrow CORS by default ## 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.