Berth vs PrivateGPT

Berth 76

v0.4.1 on 1 November 2026 and v0.4.0 on 29 October, per the repository's tags and the changelog (30). Eleven releases from v0.1.0 on 11 October to v0.4.1, plus nine release candidates (20). The changelog credits outside reporters and fixes their issues within days (azrael-229's #231 and #232, TacoTakumi's and cygnal's #189 and #196 in v0.4.0, and Windows testers' #81, #82, #103 and #104 fixed in releases tagged 20 and 21 October). Reply times and open issue counts were unreadable, and the project is 21 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). 30+20+18+0+8 = 76.

PrivateGPT 41

The last release is 1.0.1, dated 16 June 2026 in the changelog and tagged 18 June, 141 days before the check (10 of 30). No release or dated changelog entry in the 90 days since 8 August (0 of 20). Maintainers commit most days and outside pull requests merged in October, but 23 pull requests are open and the stale bot closes issues, including two security reports, with no maintainer reply (13 of 25). Homebrew tap, wheel index and Docker images ship with releases, and Anthropic SDKs work against it. No product MCP server of its own, though it connects to remote MCP servers (10 of 15). Dependabot, uv.lock, actions pinned by commit since 19 October and dependency CVE upgrades in June, July and August (8 of 10). 10 + 0 + 13 + 10 + 8 = 41.

Berth 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 at v0.4.1. 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). 12+15+8+11+5 = 51.

PrivateGPT 32

Read with the tool checklist. Authentication is off by default, and when on it is one static secret compared with the whole Authorization header, with a sample secret:key value in settings.yaml. There are no per-user keys or scopes, private-gpt serve binds 0.0.0.0 by default and CORS allows every origin, so a default start answers anyone on the network. The docs disagree on the header (Bearer, Basic or an Anthropic key) (6 of 30). Tools are chosen per request and code execution has internet off by default, but bash runs as a host subprocess with resource limits, not in a container, and nothing asks for approval (6 of 20). Web fetch, web search, documents and MCP servers bring untrusted text to the model. A public prompt-injection report (#2270) was closed by the stale bot, and no guidance was found beyond a 19 October change that bounds tool output (3 of 15). OpenTelemetry traces cover every request, model call, tool and sandbox step, off by default, with local JSON lines exporters (10 of 15). SECURITY.md since 23 July 2026 asks for private reports through GitHub or Discord. No bounty, no certification, and the June and July security reports in issues were closed by a bot with no reply (7 of 20). 6 + 6 + 3 + 10 + 7 = 32.

Berth 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 a docs page for each function (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). 6+8+17+9+14+15 = 69.

PrivateGPT 83

An OpenAPI file of 38 paths in the repository (fern/openapi/openapi.json, version 1.0.1), also served by each instance at /openapi.json, with the messages API modelled on Anthropic's (25). docs.privategpt.dev serves llms.txt, a Markdown copy of every page and a docs MCP server (10). The docs explain each endpoint group, startup and model discovery, and a compatibility table says which parts of the Claude API are supported, partial or missing, such as prompt caching and OAuth (15 of 20). Pydantic models with literal type discriminators and required fields, while settings and tool context are wide objects (12 of 15). curl examples for each guide and a /v1/messages/validate route, with error shapes not documented as a list (9 of 15). /v1 paths and a dated changelog, though releases trail main by more than four months (12 of 15). 25 + 10 + 15 + 12 + 9 + 12 = 83.

Berth 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, and since v0.4.1 a base URL per card, /cards/CARD/v1. The skill is about 12.9 KB at v0.4.1, 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). 18+12+19+15+11 = 75.

PrivateGPT 64

Read for an HTTP API an agent calls on a local instance. Token counting, top_k style retrieval settings, citation output and response byte caps let a caller size context (18 of 25). Artifact listing by namespace and retrieval filters, with no documented cursor pagination (12 of 20). FastAPI validation errors name the field and the validate route checks a request before running it, but there is no error catalogue (12 of 20). Async messages and ingestion return task ids with status, cancel and delete routes. No idempotency keys were found (10 of 20). Anthropic SDKs and Claude Code work by pointing ANTHROPIC_BASE_URL at the server, but a model server and an embedding server have to be configured first (12 of 15). 18 + 12 + 12 + 10 + 12 = 64.

Berth 61

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, 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 227 test files at v0.4.1. Whether it passes on main couldn't be read, since GitHub's pages refuse us and the earlier pass reading came from a third-party badge service, which this check doesn't use (15 of 25). The changelog credits outside reporters by issue number and fixes their reports within days (Windows reports #81, #82, #103 and #104 in v0.3.2 and v0.3.3, #189, #193, #196, #231 and #232 in v0.4.0 on 29 October). Open issue counts and reply times were unreadable (15 of 25). Semver tags, a Keep a Changelog file with dated Added, Changed and Fixed sections, and a pre-release channel. Behaviour changes are explained under Changed with what to do, such as v0.4.0 governing every card by default and how to keep the old behaviour, with no Breaking heading (13 of 15). v0.4.1, called alpha (0). 18+15+15+13+0 = 61.

PrivateGPT 70

Read with the local-software lines, since PrivateGPT runs on the owner's hardware with no hosted service behind it. Homebrew tap, uv tool install from Zylon's own wheel index (not PyPI) and Docker Hub images zylonai/private-gpt 1.0.1 of 18 June 2026. Python 3.11 only, stated in pyproject.toml, and it needs a separate OpenAI-compatible model server (17 of 20). tests.yml runs format, lint, typecheck and the tests affected by each push on main, with the full suite of 217 test files on release pull requests. We couldn't see run results because the GitHub API refused us, and issue #2441 of 1 November reports the Claude spec update workflow failing (15 of 25). GitHub shows 3 open issues of about 1,250, because a stale bot closes quiet issues, two security reports among them, with no maintainer reply (12 of 25). release-please keeps a changelog with breaking changes marked, but no release has been cut since 1.0.1 while main carries 51 commits since 8 August, among them a breaking observability change of 23 October (11 of 15). Version 1.0.1 (15). 17 + 15 + 12 + 11 + 15 = 70.