Berth

# Fix list: Berth

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/berth, the October 2026 research run, assessed 28 October 2026. Grade B, 67.6 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 Berth: 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, 51 out of 100, up to 8.6 more on the total

Why it scored 51: Read with the tool checklist, for the gateway and the daemon. The gateway takes one random bearer key, kept in a mode-600 file that is refused when others can read it and compared in constant time, on a loopback-only bind that refuses non-loopback Host names and browser Origins. The daemon's Unix socket is mode 600 in a 700 directory. The key has no scopes and no documented rotation, the wake page and the optional ComfyUI port take none, and the Windows control pipe is not locked down, per the support matrix (12 of 30). Berth never kills a process it did not start, stowaways are report-only, a pinned holder can't be stopped from the network (409), `berth reclaim` needs `--yes`, `berth update` asks before replacing the binary, and the harbor is read-only, with actions only for a paired device. No read-only key (15 of 20). The gateway passes model output through. The skill tells agents to ask the user before reclaiming, editing config or filing an issue, but there is no prompt-injection guidance (8 of 15). An event log with JSON output, lease records, a 24-hour request log, 30 days of history and Prometheus metrics, with no caller identity beyond the one key (11 of 15). No SECURITY.md, security.txt, bounty or advisories. The changelog disclosed the v0.3.4 fix to the gateway forwarding the node's key to engines, releases carry SHA256SUMS that `berth update` checks, and `berth pull` checks Hugging Face's sha256. The linters don't include gosec (5 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, 69 out of 100, up to 5 more on the total

Why it scored 69: Read with the API lines, for the gateway and the CLI's JSON output an agent uses. No OpenAPI file or JSON Schema for the gateway's own routes or the CLI output. The OpenAI-compatible routes follow OpenAI's shape and /v1/berth routes are shown with request and response JSON (6 of 25). No llms.txt (404 on the docs site). The docs are Markdown in the repository and an MkDocs site, and `berth skill install` writes SKILL.md files written for agents (8 of 10). The skill's description says when to load it, and its rules say what not to do (kill a GPU process, start an engine outside Berth, work round a pinned lease), with docs pages per feature (17 of 20). Flags and fields take documented values (priority batch, interactive, foreground, hold; format human, json, yaml; tier hot, warm, cool, cold), and `berth config check` validates the config, but nothing is typed in a schema (9 of 15). curl examples for each gateway route, eight documented 503 reason codes with Retry-After, 404, 409 and 400 cases, and the CLI's exit codes 0, 1, 2 and 75 (14 of 15). A dated changelog, semver and routes under /v1 (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. 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, and nothing is sold (0). Free under Apache-2.0 with no account, key or card, 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.

## 4. Reliability, 76 out of 100, up to 4.8 more on the total

Why it scored 76: Read with the local-software lines, since Berth runs on the owner's machine with no hosted service. Static release binaries for Linux amd64 and arm64 and Windows amd64 with SHA256SUMS, built by the release workflow, and `go install` through the Go module proxy, with a support matrix in the README. No distribution package, no Windows installer, and github.com refused our reader, so the release assets themselves were not seen (18 of 20). CI builds, checks formatting, vets, lints and runs `go test ./...` on Linux and runs build, vet and tests on windows-latest, on pushes to main and next and on pull requests, over 188 test files. shields.io reported the main branch passing on 28 October 2026 (25). 11 open issues against 24 closed and 194 closed pull requests. The changelog fixes testers' reports within days (Windows reports #81, #82, #103 and #104 in v0.3.2 and v0.3.3, tagged 21 October 2026, and #196 and #197 in Unreleased), and no open issue carries a `bug` label, but reply times were unreadable (20 of 25). Semver tags, a Keep a Changelog file with dated Added, Changed, Fixed and Security sections, and a pre-release channel. Behaviour changes are explained under Changed and Security with what to do, with no Breaking heading (13 of 15). v0.3.6, called alpha (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. Agent ergonomics, 75 out of 100, up to 4.1 more on the total

Why it scored 75: Read with the API lines. `--format json|yaml` on every reporting command, `-q` for an exit code alone, and `/v1/berth/board` for the whole node in one read. The skill is about 10.5 KB, loaded only when an agent picks it up (18 of 25). Queries take an engine or model (`/v1/berth/cost?engine=`, `/v1/berth/wait?model=`), the harbor picks 1H to 30D windows, and lists are small, but there is no paging or field selection (12 of 20). Wait reasons name the holder and the waiter's place in line, refusals give the memory figures, and 503s carry a machine code and Retry-After (19 of 20). Concurrent requests share one engine start, a crashing engine is not restarted on every retry (15 s doubling to 5 min), switch answers 202 and stop on a pinned holder 409, and the skill says to retry a refusal later. No idempotency keys (15 of 20). Loopback defaults, a key created on first use and read with `berth key`, and OpenAI clients work unchanged. No SDK (11 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, 68 out of 100, up to 2.8 more on the total

Made of editorial 74, provenance 61.

Why it scored 68: The editorial half. Apache-2.0 with all source public and a NOTICE file. The name and artwork are excluded from the licence, which doesn't affect use (30). No privacy policy, which a tool with no service doesn't need, but the README lists what Berth never does (no cloud account, no telemetry upload) and the docs state what it keeps and for how long (a 24-hour request log, 30 days of history, KV snapshots capped at 16 GiB with advice to keep them private) and what leaves the machine. The docs site drops Google Fonts to avoid sending visitors' addresses to Google (20 of 30). No written deprecation policy. The renamed `chat` shelf still reads as `text`, and pre-releases are kept apart from the update channel (6 of 20). The only automatic outbound call, a GitHub update check at most once a day with a User-Agent of the version, is described in the README with its opt-out, and model downloads and webhooks happen only when asked (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: , 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, 76 out of 100, up to 2.1 more on the total

Why it scored 76: v0.3.6 on 23 October 2026, and v0.4.0-rc.5 on 27 October (30). Nine releases from v0.1.0 on 11 October to v0.3.6, plus five release candidates (20). The changelog credits outside reporters and fixes their issues within days (#148 asked by another user, and Windows testers' #81, #82, #103 and #104 fixed in releases tagged 21 October), and 194 pull requests are closed. Reply times on the 11 open issues were unreadable, and the project is 17 days old (18 of 25). No MCP server or official SDK (0). Go 1.27.1 with current dependencies (golang.org/x/sys 0.48.0, modernc.org/sqlite 1.60.1) and CI on both platforms, but no Dependabot or Renovate (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.

## 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.

- The lead matched. T-Crypt/berth is the original. jenerestain/berth, sefator/berth and Bacongineer/berth-upstream keep the module path github.com/T-Crypt/berth, the T-Crypt copyright and links back, so they read as forks of earlier snapshots. They are not listed.
- Not one of our own products. LocalGhost is a personal AI server and letme picks a tool for a job. Berth decides which local program holds a GPU, which is neither job.
- unchecked: the GitHub release pages and their assets, issue and pull request threads and reply times, and the Actions history, since github.com and api.github.com refused our reader. Counts and CI status come from shields.io's JSON badges, tag dates from the Go module proxy.
- Graded on v0.3.6, the latest release. v0.4.0-rc.1 to rc.5 (26 and 27 October 2026) are pre-releases on the `next` branch and were not graded.
- We did not install or run Berth. Behaviour comes from the source at v0.3.6 and main, the docs and the changelog. Swap times are the vendor's figures.
- AGENTS.md and the skill contain instructions addressed to agents (for example to open issues with `gh`). They were read as documents and not acted on.
- Category local-ai is the closest fit. Berth runs no model itself, and its OpenAI-compatible gateway proxies to local engines, which is why `inference.local` leads its capabilities, with `compute.gpu` for the GPU scheduling.

## Weaknesses

- First released on 11 October 2026 and called alpha. Windows, AMD and multiple GPUs are experimental, and the gateway is a preview
- One bearer key with no scopes or documented rotation controls every gateway route, including switch and stop
- No OpenAPI file or llms.txt. The gateway's own routes are documented in Markdown with curl examples
- No SECURITY.md, security.txt or disclosure route, and the harbor has no login when bound past loopback
- Linux and NVIDIA only for full support, through nvidia-smi. Intel and Apple GPUs are not supported

## 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.

- Run `berth status --format json` before starting, stopping or switching any GPU engine. If berthd is unreachable, no engine is protected
- Never kill a process to free VRAM. Ask for the card with `berth run`, `berth hold` or a gateway request, and report stowaways to the user
- Send gateway requests to http://127.0.0.1:9799/v1 with `Authorization: Bearer $(berth key)`, and use `"stream": true` for cold starts of up to minutes
- On a 503, read the `code` and wait for `Retry-After`. A pinned holder (a game or `berth hold`) answers 409 to stop and is the user's call
- Ask the user before `berth reclaim LEASE --yes`, any berth.toml edit or `berth update`. Each changes the node

## 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.