# Berth (slim)

> Berth is an open-source Go daemon and CLI that leases one GPU to local inference engines, image pipelines and games in turn, with a priority queue, measured VRAM handoffs and a preview OpenAI-compatible gateway on loopback.

- Full: https://www.anchorterminal.com/tools/berth.md (~7,950 tokens) · this version ~1,830 tokens · JSON https://www.anchorterminal.com/tools/berth.json · canonical https://www.anchorterminal.com/tools/berth
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-28 · pricing https://www.anchorterminal.com/tools/berth#pricing · reviews https://www.anchorterminal.com/tools/berth#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 67.6/100 · rank #283 of 1218 · #2 in Local AI · not agent-ready · confidence medium**


Assessment: Berth coordinates one NVIDIA card between engines, ComfyUI and games, grants a lease only after the driver shows memory free, and gives agents a skill, JSON output and a keyed loopback gateway. It is 17 days old, alpha and maintained by one developer, with Windows and AMD experimental and no machine-readable API spec.

## Facts

- Kind: HTTP API · vendor: Trevin Tindall (T-Crypt on GitHub) · category: Local AI · legal entity: Trevin Tindall (T-Crypt on GitHub) · provenance 61/100
- Local only (HTTP): go `github.com/T-Crypt/berth`
- Auth: API key · pricing: Free · x402: no · licence: Apache-2.0. The Berth name, the Pilot mascot and the artwork in assets/ are excluded from the licence, per the NOTICE file
- Probe metrics: not measured yet (probes haven't run)
- Surface listed: Berth v0.3.6, its CLI with JSON output, the agent skill and the preview gateway, read from the repository at main and tag v0.3.6, the docs and the changelog on 28 October 2026. Nothing was installed or run
- Interfaces: The `berth` CLI (daemon, run, hold, stop, resume, reclaim, status, events, models, pull, gateway, harbor, skill and more), the daemon's Unix socket (NDJSON), the harbor web UI on 127.0.0.1:9798, Prometheus /metrics on loopback and the gateway on 127.0.0.1:9799. No MCP server
- Gateway routes: OpenAI-compatible /v1/chat/completions, /v1/completions, /v1/embeddings and a merged /v1/models, ComfyUI's /prompt, /queue, /history, /view and /ws, and Berth's own /v1/berth/board, switch, stop, wait, cost and plan. Preview
- Leasing: Priorities hold, foreground, interactive and batch. The holder finishes its reply or image job, is unloaded (its own unload, then SIGTERM, then SIGKILL for processes Berth started) and the next lease waits for the driver to show the memory free. Parking in RAM for ComfyUI, vLLM, SGLang and engines on the park protocol
- Engines read: llama.cpp, llama-swap, Ollama, vLLM, SGLang, ComfyUI, Kokoro, LM Studio, KoboldCpp, TabbyAPI, LocalAI and any OpenAI-compatible or custom server named in berth.toml
- Platforms: Linux with NVIDIA (supported). Windows 10/11 with NVIDIA, Linux AMD RDNA3, RDNA4 and Phoenix APUs, and one berthd per card on multi-GPU machines (experimental). Release binaries for Linux amd64 and arm64 and Windows amd64
- Network: An update check to api.github.com at most once every 24 hours (off with `[update] check = false`), model downloads from Hugging Face and CivitAI on `berth pull`, probes of local engine ports, and opt-in webhooks. No telemetry upload
- Data kept: Leases and events, a 24-hour request log, 30 days of telemetry history in SQLite under ~/.local/state/berth, and optional llama.cpp KV snapshots, capped at 16 GiB by default, which hold conversation prompts
- Vendor benchmarks: On the author's RTX 4090 with 62.5 GiB of RAM (23 October 2026), a 21 GB chat model from the page cache in 5.6 s and ComfyUI to a 27B chat model in 13.6 to 16.3 s, request to reply. Vendor figures
- Releases: v0.1.0 (11 October 2026) to v0.3.6 (23 October), nine releases, then v0.4.0-rc.1 to rc.5 on 26 and 27 October as pre-releases on the `next` branch
- Repository activity: 245 stars, 22 forks, 11 open and 24 closed issues, 2 open and 194 closed pull requests and 3 contributors, per shields.io's reading of GitHub on 28 October 2026

- Scores: Reliability 76, Performance pending, Schema & documentation 69, Agent ergonomics 75, Security & auth 51, Payments & pricing 60, Task success pending, Maintenance & community 76, Transparency & trust 68 · total over the 7 assessed categories
- Why: Reliability, Read with the local-software lines, since Berth runs on the owner's machine with no hosted service. · Schema & documentation, Read with the API lines, for the gateway and the CLI's JSON output an agent uses. · Agent ergonomics, Read with the API lines. · Security & auth, Read with the tool checklist, for the gateway and the daemon. · Payments & pricing, Read with the self-hosted rule. · Maintenance & community, v0.3.6 on 23 October 2026, and v0.4.0-rc.5 on 27 October (30). · Transparency & trust, The editorial half.
- Sources: 19, open questions: 7, both in the full twin
- Capabilities: inference.local, compute.gpu
- JSON: https://www.anchorterminal.com/api/v1/tools/berth.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/berth.svg` or a link to https://www.anchorterminal.com/tools/berth from a page under github.com/t-crypt, or the README of github.com/T-Crypt/berth, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.
- Support: https://github.com/T-Crypt/berth/issues; https://t-crypt.github.io/berth/; more in the full twin

## Before you call it

1. Run `berth status --format json` before starting, stopping or switching any GPU engine. If berthd is unreachable, no engine is protected
2. 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
3. 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
4. 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
5. Ask the user before `berth reclaim LEASE --yes`, any berth.toml edit or `berth update`. Each changes the node

## Connect

```bash
go install github.com/T-Crypt/berth/cmd/berth@latest      # or a release binary checked against SHA256SUMS
berth service install --harbor
berth status
berth run --as llama --priority foreground -- llama-server -m model.gguf --port 8080
```

```bash
curl -s http://127.0.0.1:9799/v1/berth/board -H "Authorization: Bearer $(berth key)"
```

```bash
berth skill install   # writes the Berth and ComfyUI skills, including to ~/.claude/skills
```
Full config and headless snippets are in the full page. Through letme (picks today, calling later): https://letme.dev/berth

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| LocalAI | B | 68.5 | inference.local | https://www.anchorterminal.com/tools/localai.min.md |
| Nebius AI Cloud | B | 67.3 | compute.gpu | https://www.anchorterminal.com/tools/nebius-ai-cloud.min.md |
| Baseten | B | 66.7 | compute.gpu | https://www.anchorterminal.com/tools/baseten.min.md |
| Hugging Face Inference Endpoints | B | 64.7 | compute.gpu | https://www.anchorterminal.com/tools/hugging-face-inference-endpoints.min.md |
| Lemonade | B | 63.9 | inference.local | https://www.anchorterminal.com/tools/lemonade.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)