Harbor: download ældre versioner

Misc
- Harbor AppImages now include a portable icon and launch correctly for ordinary users.
- The Harbor desktop app opens reliably on Linux.
- The README now identifies Harbor correctly in the AppImage catalog.
Full Changelog: https://github.com/av/harbor/compare/v0.5.16...v0.5.17

rummage

Local multimodal semantic search over your files: one program serves a web UI and an HTTP API, watches the folders you choose, and indexes text, images, PDFs, audio and video with Google DeepMind's EmbeddingGemma 2 on llama.cpp, using the CPU, NVIDIA (CUDA, including CDI) or AMD (Vulkan by default, ROCm optional).
harbor up rummage
Harbor (Agent Evaluations)

Laude Institute's Harbor Framework for evaluating agents in local Docker tasks, from the creators of Terminal-Bench, available under the harbor handle.
harbor up harbor
harbor harbor --help
Misc
- harbor dev seed-cdi keeps CDI overlay files that are maintained by hand (without the generated header), so a service can select its GPU image under CDI too.
Full Changelog: https://github.com/av/harbor/compare/v0.5.15...v0.5.16

Misc
- harbor models pull now downloads only the files you name. harbor models pull --source hf <repo> <file...> used to ignore the filenames and fetch the whole repository; the filenames are now passed to the Hugging Face CLI. A pull with filenames and no --source also goes through the Hugging Face CLI, including for GGUF repositories. Pulls with only a repository name work as before. Thanks to @cortezbuilds for the report and the fix (#290, #291).
harbor models pull --source hf hf-internal-testing/tiny-random-gpt2 config.json
Full Changelog: https://github.com/av/harbor/compare/v0.5.14...v0.5.15

Latitude

Self-hosted agent observability that collects traces, groups failures into signals, and runs evaluations, using the official single-host images with local datastores and a Mailpit inbox for sign-in.
harbor up latitude
AutoGen Studio

Low-code web UI for prototyping agents and multi-agent workflows, built from the published autogenstudio package and wired to Ollama.
harbor up autogenstudio
LobeHub

Database-backed successor to LobeChat with accounts, persistent chats in PostgreSQL, knowledge files in S3-compatible storage, and Ollama chat and embeddings. The existing lobechat handle stays as the browser-storage variant.
harbor up lobehub
R2R

Document ingestion, semantic search and RAG answers with a dashboard, persistent pgvector database and graph clustering, backed by local Ollama models.
harbor up r2r
Caddy
Reverse proxy with optional automatic HTTPS, started with a local-only health route and a user-selectable Caddyfile for the Harbor services you want to expose.
harbor up caddy
RAGFlow

Document-focused RAG engine with parsing, datasets, hybrid retrieval and chat/agent workflows, packaged as upstream's CPU stack with persistent Elasticsearch, MySQL, S3-compatible storage and Valkey.
harbor up ragflow
ClaraVerse

Private AI workspace for chat, agent teams and visual workflows, running the official image with persistent MySQL, MongoDB and Redis and Ollama model discovery.
harbor up claraverse
Qwen Code

Coding agent with a terminal CLI and a browser Web Shell, running the official image with a persistent project workspace and Ollama integration.
harbor up qwencode
Misc
- Harbor can now run from a disposable Docker container via scripts/harbor-docker.sh, with the workspace and caches at their host paths and the CLI acting as your user; see Installing Harbor.
- Boost and Open WebUI can use an Ollama running outside Harbor, including one on the host, without starting the ollama service.
- Opt-in command wrappers let you call service CLIs such as aider directly from the shell, with collision protection and sync.
- Deno routines can trust a custom CA certificate and use an alternate Docker network for proxied or restricted networks.
- The Ollama ROCm image tag can be pinned instead of tracking the rolling rocm tag.
- harbor doctor no longer falsely reports ROCm from an lsmod pipe failure and detects built-in kernel modules.
- Perplexica initializes correctly when opened from another machine on the LAN and its backend URL is configurable.
- AnythingLLM keeps its chosen chat and embedding models across harbor restart.
- OpenCode keeps manual and inline config and can discover models from an authenticated LiteLLM.
- The Harbor Agent image builds again on the current Webtop base.
- The Harbor App reads and writes profiles when the home directory is dot-prefixed.
- The last case-sensitive broken link in the docs is fixed.
Full Changelog: https://github.com/av/harbor/compare/v0.5.13...v0.5.14

Misc
- harbor models pull no longer prints "Download completed successfully" when the download actually failed. A failed Hugging Face resolve is now reported as an error, and a cache your user cannot write to points you at harbor fixfs.
- The PrismML backend documents both Bonsai 2 quants with measured numbers: PTQ1_0 (5.95 GB) and PQ2_0 (7.21 GB) answered 19/20 of the same prompts correctly, but PQ2_0 generates about 25% faster.
- The PrismML docs warn that raising -np above one slot badly degrades answer quality on ROCm; Harbor's default of a single slot is unaffected.
Full Changelog: https://github.com/av/harbor/compare/v0.5.12...v0.5.13

Misc
- Long Halogen thinking turns can finish instead of stopping at 8192 tokens.
- The Halogen engine stays up through long silent decodes for 900 seconds.
- Halogen 0.6.3 sizes the KV pool so the lookup-table file cache still has room beside Open WebUI.
Full Changelog: https://github.com/av/harbor/compare/v0.5.10...v0.5.11

Misc
- n8n first-time setup includes the Assistant code sandbox so new instances can finish the GUI wizard.
Full Changelog: https://github.com/av/harbor/compare/v0.5.9...v0.5.10

Misc
- Hugging Face downloads run as your configured host user and recover root-owned caches automatically after upgrading.
- The HF permissions fix applies to existing downloader images without a manual rebuild (fixes #262).
Full Changelog: https://github.com/av/harbor/compare/v0.5.8...v0.5.9

Halogen
Halogen Flash Server serves Qwen3.8 through an OpenAI-compatible API on AMD Strix Halo, with an Open WebUI integration.
harbor up halogen
Misc
- Hugging Face downloads now work when Docker created the cache directory as root, while new files remain owned by your configured host user (fixes #262).
After upgrading, run harbor build hf once to rebuild an existing downloader image with the fix.
Full Changelog: https://github.com/av/harbor/compare/v0.5.7...v0.5.8

🦭 TL;DR
A whole menagerie in one release. 🦉 Chandra 2 joins the backend flock as a vLLM-served OCR model that reads documents and images into Markdown, HTML or JSON. 🐙 Boost grows a new arm with the codemode module, where a model writes one Python program that calls every Boost tool as a plain function, and harbor launch --codemode wires it into any launched CLI. 🦔 harbor down <service> and harbor restart <service> now touch only the named services, 🐜 the ollama-init sidecar shrinks from the multi-GB Ollama image to Alpine, 🐿️ services can ship their own default.env, and 🐝 harbor dev docs --check keeps generated docs honest.
🦉 Chandra
Chandra 2 OCR by Datalab, served with vLLM behind an OpenAI-compatible API: turns documents and images into Markdown, HTML or JSON with tables, forms, math and reading order intact. Needs an NVIDIA GPU; 16GB is enough.
harbor up chandra nvidia
🐙 Boost codemode
Replaces the whole Boost tool catalog with a single execute_code tool. The model writes one Python program that calls Boost's tools as functions; Boost runs it in a sandboxed subprocess runner and round-trips every call to the real tool, with caps on timeout, output size and number of tool calls.
harbor launch --codemode
harbor boost modules add codemode
🐾 Misc
- 🦔 harbor down <service> and harbor restart <service> are scoped to the named services; other running services are left untouched.
- 🐜 Slim ollama-init sidecar: runs on alpine:3.20 instead of the multi-GB Ollama image and streams pulls through the server's /api/pull API.
- 🐿️ Per-service defaults: services/<name>/default.env files are loaded with the profile.
- 🦫 Multi-source .env repair is atomic and duplicate-free.
- 🐝 harbor dev docs --check flags stale generated docs, and the release pipeline builds Boost before regenerating its docs.
- 🐘 Chandra defaults tuned for 16GB GPUs.
Full Changelog: https://github.com/av/harbor/compare/v0.5.6...v0.5.7

TL;DR
Six new services — Text Embeddings Inference, LightRAG, Paperless, Paperless-GPT, Whishper, and Linkwarden — each with cross-service integrations, Traefik routes, host-owned workspaces, and runnable documentation.
Text Embeddings Inference
Hugging Face's embedding and reranking server, wired into Open WebUI, LibreChat, and AnythingLLM.
harbor up tei
LightRAG
Graph-based retrieval augmented generation with a web UI, running against Ollama or llama.cpp with TEI embeddings, and available as a model in Open WebUI.
harbor up lightrag
Paperless
Document management system that scans, indexes, and archives your paper documents.
harbor up paperless
Paperless-GPT
LLM-generated titles, tags, and OCR for Paperless documents.
harbor up paperless-gpt
Whishper
Self-hosted audio transcription, subtitle editing, and translation built on Whisper and LibreTranslate.
harbor up whishper
Linkwarden
Bookmark manager that archives pages and tags them automatically with a local LLM.
harbor up linkwarden
Misc
- Traefik routes for all six new services, so each is reachable on its own hostname.
- Files the new services write to your workspace stay owned by your host user.
- LightRAG answers complete on AMD hosts via a ROCm cross-file that keeps the query model on CPU.
- TEI accepts whole-document ingest from AnythingLLM.
- Linkwarden and Paperless logins stay on their Traefik hostnames instead of bouncing to localhost.
- Whishper repairs workspace ownership on start and mirrors worker output to docker logs.
Full Changelog: https://github.com/av/harbor/compare/v0.5.5...v0.5.6

Misc
A hardening release: a runnable integration suite now exercises 40 Harbor services end-to-end, and everything it caught is fixed.
- New tests/services-integration.sh runs the whole service integration suite (grouped, with --groups/--list), backed by a written spec in tests/services-integration.md.
- First-boot failures fixed across many services: SearXNG (stale settings crash-loop), Langflow and AnythingLLM (storage permissions), LibreChat (RAG database race and a broken Ollama endpoint URL), LibreTranslate, Plandex (install source, server image, port, postgres pin), webtop builds, fabric, cmdh, and metamcp headless start.
- chat-ui works with Harbor backends again on current upstream versions.
- LDR's SearXNG search integration and persistent data storage now work; Presenton reaches Ollama and no longer requires login setup.
- Bifrost restarts cleanly with a persisted config; optillm is reachable on its published port; mcp-inspector no longer spawns runaway processes.
- LiteLLM gains ready-made llama.cpp and optillm routing configs.
- harbor up accepts cross-file service selectors like harbor up mcpo mcp-server-time.
- Boost now returns the backend's real HTTP status and error body instead of an empty 200 when a request is rejected, and logs handled backend errors as a single concise line.
- harbor pull rejects unknown service names with a clear error instead of starting Ollama.
- ROCm GPU variants (llama.cpp, Ollama, LocalAI, Lemonade) verified with real inference on AMD hardware, with the runner detecting ROCm hosts automatically.
Full Changelog: https://github.com/av/harbor/compare/v0.5.3...v0.5.4

Misc
- Harbor Boost gains agentic modules: quickhop and deephop for web research, autocheck for deliverable audits, and diffscope for scoped file-change checks, with caveman and ponytail as response style modules.
- harbor launch --workflow routes coding agents through Boost module chains.
- harbor launch supports the grok CLI as a host tool.
- Faster Hugging Face downloads via hf_transfer.
- harbor hf cache is renamed to harbor hf cachedir, freeing hf cache to pass through to the official HF CLI.
- HF cache scans now rescue models the hub library misses, keep listing models when the cache scan throws, and llamacpp models without a slash in their ID can be removed.
- Boost read_url now validates every redirect hop, closing an SSRF vector in web research modules.
- Boost autocheck keeps failed or spoofed page reads out of its evidence, and the /caveman and /ponytail style commands only trigger in their own module.
- Boost token counting no longer stalls on multi-megabyte pathological inputs.
- harbor pull and other compose-backed commands now fail with a clear error when compose resolution fails, instead of running a bare docker subcommand.
- harbor up a b -t now tails logs for all requested services, not just the first.
- Assorted CLI fixes: doctor --check ordering, up --open target, and ui.main resolution for open/url/tunnel/qr.
- harbor pull handles mixed service/model/flag arguments correctly, routing flags like --no-defaults to the compose resolver instead of misreading them as model names.
- Host Deno runs no longer show update-check nags or send telemetry.
- The container test runner auto-runs the install suite when needed and works on Alpine, Arch, and Rocky; compose files can waive lint rules via x-harbor-lint-ignore.
Breaking: built-in Boost workflow presets and the keel/sightline modules were removed; the former caveman/ponytail research workflows are now quickhop/deephop.
Full Changelog: https://github.com/av/harbor/compare/v0.5.2...v0.5.3

TL;DR
Four new backends - DMR (Docker Model Runner), MLX, oMLX, and Daytona sandbox - with cross-service integrations rolled out to 26 services each. llamacpp replaces Ollama as the default backend (harbor defaults add ollama to restore the old behavior). The Tauri app now handles the entire install workflow in-app. New CLI features: tab completion, port conflict detection, smart pull routing, custom volume mounts, fuzzy service suggestions, and harbor skills for AI agent discovery.
llamacpp default backend (breaking change)
llamacpp is now the default backend, replacing Ollama. Running harbor up with no arguments starts llamacpp instead of Ollama. Existing users who rely on harbor up launching Ollama should explicitly run harbor up ollama or set their default via harbor defaults add ollama. Cross-service integrations for llamacpp have been rolled out to 38 services.
Agent Skills
New harbor skills command exposes agent-facing CLI documentation, letting AI coding agents discover and use Harbor commands without manual prompting.
harbor skills # list available skills
harbor skills get <name> # show a skill's content
In-App Install
The Tauri app now ships a full guided setup flow — Docker detection, dependency installation, platform-specific package management, and CLI bootstrapping all happen inside the app. Covers macOS, WSL, Fedora, openSUSE, and minimal environments with cancel-between-steps, error recovery, and disk space checks.
DMR (Docker Model Runner)
Run models through Docker Desktop's built-in Model Runner with OpenAI-compatible API, auto-bootstrapped on harbor up.
harbor up dmr
MLX
Apple Silicon native inference via mlx-lm, with auto-bootstrap and host component setup on first run.
harbor up mlx
oMLX
OpenAI-compatible MLX inference server for Apple Silicon.
harbor up omlx
Daytona
Self-hosted sandbox environment service for AI agent workloads.
harbor up daytona
Misc
- Tab completion for bash, zsh, and fish shells.
- Port conflict pre-detection warns before harbor up fails.
- Smart pull routing automatically directs model downloads to the right backend.
- Custom volume mounts via harbor volumes CLI.
- Fuzzy service name suggestions when you mistype a service handle.
- MiniMax default model upgraded to M3.
- Comprehensive integration test suites added for CLI, config, and service lifecycle.
Full Changelog: https://github.com/av/harbor/compare/v0.4.19...v0.5.0

Changes
https://github.com/user-attachments/assets/e4897391-c5a8-4391-93c3-9f8b76155f11
A new launch command to spawn pre-configured local agentic coding tools on host pointing to inference backends in Harbor.
# Uses the first reachable running backend and discovered model
harbor launch codex

# Explicit backend and model selection; launch options go before the tool name
harbor launch --backend ollama --model qwen3.5:4b codex
harbor launch --backend ollama --model qwen3.5:4b mi
harbor launch --backend llamacpp --model Qwen3.5-4B opencode
harbor launch --backend ollama --model qwen3.5:4b copilot -p "explain this repo"

# Route through Harbor Boost with web search and URL-reading tools
harbor launch --web --backend ollama --model qwen3.5:4b codex

# Claude Code uses Harbor's Ollama backend through Ollama's Anthropic-compatible API
harbor launch --backend ollama --model qwen3.5:4b claude -p "explain this repo"

# Print the computed launch config without starting the tool
harbor launch --config opencode
harbor launch --config openclaw
- Anthropic-compatible Messages API and OpenAI Responses API compat layers for Boost
- harbor launch command for starting services with host tool adapters
- ik_llama.cpp service
- Boost workflows and tools module
- Adaptive thinking, annotations, rate limit forwarding, SSE keep-alive in compat layers
- Extensive Boost hardening: streaming fixes, SDK compat audit, request validation
- new guides, docs screenshots optimised
Full Changelog: https://github.com/av/harbor/compare/v0.4.18...v0.4.19

Open Design
Local-first design artifact workspace with BYOK model providers, design systems, decks, prototypes, and Harbor backend integrations.
harbor up open-design
Voicebox
Local-first AI voice studio for voice cloning, preset voices, speech generation, dictation, transcription, and agent speech.
harbor up voicebox
Misc
- harbor how now uses mi instead of cmdh.
- Open Design includes Ollama, llama.cpp, and vLLM integration presets.
- Voicebox includes NVIDIA, CDI, and ROCm overlays with Harbor-managed data, generations, and Hugging Face model cache directories.
Full Changelog: https://github.com/av/harbor/compare/v0.4.17...v0.4.18

Needle
Needle adds a Harbor-built OpenAI-compatible tool-calling backend for Cactus Compute's small function-calling model.
harbor up needle
npcsh
npcsh adds an agentic shell/API service with local backend integrations and frontend/gateway wiring.
harbor up npcsh llamacpp
Full Changelog: https://github.com/av/harbor/compare/v0.4.16...v0.4.17

ML Intern
Hugging Face ML engineering agent for researching papers, inspecting Hugging Face resources, training models, and shipping ML-related code with local or cloud LLM backends.
harbor up ml-intern ollama --open
facts
Fact-driven development CLI for project specs that agents can verify and implement from inside Harbor.
harbor facts list
mi
Minimal agentic coding CLI that runs against OpenAI-compatible local or cloud model APIs.
harbor mi -p "inspect this project and summarize it"
Misc
- ML Intern llama.cpp auto mode selects ranked text/code models from the advertised router catalog and rejects unsuitable non-chat models.
- ML Intern can use GitHub token override env and reports unhealthy LLM backends with HTTP 503.
- Harbor config search now matches hyphenated service handles such as ml-intern against normalized config keys.
Full Changelog: https://github.com/av/harbor/compare/v0.4.15...v0.4.16

Lemonade
AMD Lemonade inference server for ROCm GPUs.
harbor up lemonade
Tokscale
Token-level autoscaling proxy for LLM backends.
harbor up tokscale
Misc
- TabbyAPI switched to official image with updated default model and hfdownloader fix
- fix(profile): validate remotely downloaded profiles to prevent command injection (CWE-78) by @sebastiondev in https://github.com/av/harbor/pull/236
- Lint exclusions for vendored third-party caches
Full Changelog: https://github.com/av/harbor/compare/v0.4.14...v0.4.15

Bifrost
Fast OpenAI-compatible AI gateway with provider routing, failover, observability, and Harbor backend bootstrapping.
harbor up bifrost
SurfSense
Private NotebookLM-style research workspace with RAG, citations, browser capture, reports, and podcast generation.
harbor up surfsense
Misc
- Added Harbor CLI sidecar integrations for local agent/editor workflows.
- Added Open WebUI and Harbor Boost integration paths through Bifrost.
- Fixed Unsloth Studio bootstrap portability and Docker Compose run argument passthrough.
Full Changelog: https://github.com/av/harbor/compare/v0.4.13...v0.4.14

Unsloth Studio
No-code fine-tuning UI for the Unsloth library, with a zero-click API key bootstrap that registers Studio as an OpenAI-compatible backend across other Harbor services.
harbor up unsloth-studio
Cross-integrations ready out of the box: Open WebUI, Boost, Aider, OpenCode, Hermes, OpenClaw.
Misc
- Workspace bind mounts on 17 services now stay owned by your host user — recoverable without sudo.
- harbor up no longer errors with invalid spec when HARBOR_HF_CACHE is empty.
- unsloth notebooks now see models pulled by other Harbor services through the shared HF cache.
- langfuse analytics and queue state persist across harbor down/up.
- Aider connects to several cross-integration backends that were previously broken by a bad config mount path.
Full Changelog: https://github.com/av/harbor/compare/v0.4.12...v0.4.13

Beszel
Lightweight server monitoring with Docker stats, pre-provisioned to track the Harbor host on first run.
harbor up beszel
DBHub
Universal database gateway exposed as MCP, pre-connected to Open WebUI via native MCP for zero-click chat-with-your-database.
harbor up dbhub
Misc
- Container-based test matrix runner with per-row HARBOR_HOME isolation, env-file robustness, and a git-reset hook (harbor dev test).
- New 3-pass source linter combining shellcheck, custom HARBORxxx rules, and compose validation (harbor dev lint).
- Manual integration test rewrite focused on end-user outcomes for Beszel and DBHub.
- Hardening fixes: host-user permissions, key validation, and sibling teardown across services.
- Cleared shellcheck warnings flagged by the new lint workflow.
Full Changelog: https://github.com/av/harbor/compare/v0.4.10...v0.4.12

Misc
- Fixed SearXNG settings.yml not enabling JSON format, causing WebUI web search to return 403.
- Fixed SearXNG workspace pointing away from the service template directory.
- Fixed seed-cdi script incorrectly processing .build.yml variant files.
Full Changelog: https://github.com/av/harbor/compare/v0.4.9...v0.4.10

Misc
- New Boost module analogical: generates analogies to ground LLM responses.
- New Boost module deaf: humorously misinterprets user queries and responds to the warped version.
- Open WebUI now defaults to native function calling mode.
- Fixed CLI validation to accept variant-backed services.
- Fixed deaf module to use a clean intermediate LLM instance for misheard calls.
Full Changelog: https://github.com/av/harbor/compare/v0.4.8...v0.4.9

Solo CLI
Local-first AI agent for automating tasks and deploying Physical AI on edge devices, with support for Ollama, vLLM, and llama.cpp backends.
harbor build solo
harbor up solo
ROS MCP Server
MCP bridge that lets LLMs interact with Robot Operating System (ROS) nodes, topics, and services via rosbridge.
harbor build ros-mcp-server
harbor up ros-mcp-server
Misc
- feat: add MiniMax as first-class cloud LLM provider (M2.7 default) by @octo-patch in https://github.com/av/harbor/pull/221
- feat: add solo-cli and ros-mcp-server services by @ddiddi in https://github.com/av/harbor/pull/227
- Remove deprecated VLLM_ATTENTION_BACKEND env var and docs by @genevera in https://github.com/av/harbor/pull/228
- Comprehensive documentation correctness pass across all services and backends.
- Fixed multiple CLI edge-cases: TTY handling, stats hanging, config update data loss, and command bubbling.
- Fixed restart to preserve all active services; fixed SillyTavern init script and server_urls merging.
- Added Tauri E2E automated testing with WebdriverIO.
New Contributors
- @octo-patch made their first contribution in https://github.com/av/harbor/pull/221
- @ddiddi made their first contribution in https://github.com/av/harbor/pull/227
Full Changelog: https://github.com/av/harbor/compare/v0.4.7...v0.4.8

Hermes Agent
Autonomous AI agent by Nous Research with persistent memory, extensible skills, multi-platform gateway, and OpenAI-compatible API.
harbor up hermes
Misc
- llamacpp can now be built from source with harbor llamacpp build on (CUDA and ROCm supported).
- Rewrote llamacpp documentation with updated model management workflow.
- Added rocm and build as recognized service capabilities.
Full Changelog: https://github.com/av/harbor/compare/v0.4.6...v0.4.7

SillyTavern
Feature-rich, locally-installed web UI for interacting with LLMs with support for multiple backends, personas, lorebooks, and extensions.
harbor up sillytavern
Misc
- Fixed llama.cpp cache location in a container.
- Added Jupyter workspace to fixfs for proper permissions handling.
- Removed deprecated configuration files for uWSGI and tool settings.
Full Changelog: https://github.com/av/harbor/compare/v0.4.5...v0.4.6

Harbor App Updates
This release introduces multiple QoL improvements in the Harbor desktop app.
- You can pin favorite services, service cards are cleaner
- UI for harbor models - download llama.cpp, HuggingFace and Ollama models through the app
- Globally available terminal pane with a global (Ctrl + `) shortcut
- Service details page allows to open and tail service logs
- Application sidebar shows shortcuts to currently running active services
- new harbor log aliaas for harbor logs or harbor l
- Fix: concurrent harbor ls during harbor down causing overly high CPU usage
- Fix: model pulling logic for llama.cpp now uses run_llamacpp_pull.
Full Changelog: https://github.com/av/harbor/compare/v0.4.4...v0.4.5

Open Terminal
Lightweight remote shell and file-management API for AI agents — works with Open WebUI out of the box.
harbor up openterminal
harbor models
You can now manage your llama.cpp, HuggingFace and Ollama models in a single concise CLI.
$ ▼ harbor models ls
SOURCE MODEL SIZE DETAILS
ollama qwen3.5:35b 23.9 GB qwen35moe 36.0B Q4_K_M
hf hexgrad/Kokoro-82M 358 MB
hf Systran/faster-distil-whisper-large-v3 1.5 GB
llamacpp unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF:Q4_0 45.3 GB Q4_0

# Use programmatically with jq and other tools
harbor models ls --json

# Pull Ollama models or HF repos
harbor models pull qwen3:8b
harbor models pull bartowski/Llama-3.2-1B-Instruct-GGUF

# Use same ID you can see in `ls` for removing the models
harbor models rm qwen3:8b
Misc
- Add integration tests and mock OpenAI service for CI.
- Fix workspace paths in default.env to point to services directory.
- Fix JSON merger output path in Open WebUI start script.
- Fabric docs updates and CLI fixes.
Full Changelog: https://github.com/av/harbor/compare/v0.4.3...v0.4.4

Misc
- Fixing incorrect CLI versioning
Full Changelog: https://github.com/av/harbor/compare/v0.4.0...v0.4.1

Misc
- Renamed Moltbot to OpenClaw across compose files, docs, and assets.
Full Changelog: https://github.com/av/harbor/compare/v0.3.40...v0.3.41

Misc
- Rename Clawdbot to Moltbot (handle changed to moltbot).
Full Changelog: https://github.com/av/harbor/compare/v0.3.39...v0.3.40

Clawdbot
Personal AI assistant that connects to your chat channels through a self-hosted gateway.
harbor up clawdbot
Misc
- Added support for llamacpp model pulls via harbor pull <model>, where <model> is same model ID as used by llama.cpp for HuggingFace
Full Changelog: https://github.com/av/harbor/compare/v0.3.38...v0.3.39

OpenCode
AI-powered coding assistant with server API, terminal UI, and multi-provider LLM support.
harbor up opencode
Misc
- TypeScript-based compose transforms for dynamic configuration
- Update PostgreSQL data volume path in litellm compose file by @genevera in https://github.com/av/harbor/pull/208
- Service logos added to App sidebar
- LiteLLM PostgreSQL pinned to v18 for stability
- Routines path discovery fixes
- Documentation updates for router mode in llamacpp
Full Changelog: https://github.com/av/harbor/compare/v0.3.37...v0.3.38

Misc
- Added logprobs workflow module for Harbor Boost with interactive HTML visualization.
- Updated CLI Reference documentation to reflect harbor.sh commands.
- Fixed ask harbor link in textclip badges.
- landing service for Harbor's future landing page
Full Changelog: https://github.com/av/harbor/compare/v0.3.35...v0.3.37

Photoprism
AI-powered photo management app with face recognition, image classification, and automatic organization.
harbor up photoprism
Khoj
Open-source personal AI "second brain" with chat, semantic search, and RAG for your documents.
harbor up khoj
Misc
- Better initial setup for Omarchy: NVIDIA Container Toolkit installation instructions, Docker Compose version detection for modern non-semver versions
Full Changelog: https://github.com/av/harbor/compare/v0.3.34...v0.3.35

DeerFlow
DeerFlow is a community-driven deep research framework that combines LLMs with web search, web crawling, and multi-agent workflows to generate comprehensive research reports.
harbor up deerflow
ActivePieces
activepieces is an open-source workflow automation tool that allows you to connect apps and automate workflows with AI capabilities. It provides a visual workflow builder similar to Zapier or n8n, with 200+ app connectors and built-in AI features.
harbor up activepieces
Full Changelog: https://github.com/av/harbor/compare/v0.3.33...v0.3.34

AstrBot
Probably the easiest way to bring your local LLM into Discord/Slack and other messengers.
harbor up astrbot
Full Changelog: https://github.com/av/harbor/compare/v0.3.32...v0.3.33

Open Notebook
Open Notebook is an AI-powered research and note-taking platform that helps you organize knowledge with intelligent assistance. It features multi-provider LLM support (OpenAI, Anthropic, Google, Groq, Ollama), automatic content extraction, RAG-based search, podcast generation, and comprehensive content transformations.
harbor up opennotebook
Full Changelog: https://github.com/av/harbor/compare/v0.3.29...v0.3.30

Misc
- fixed Harbor App not showing core services with a hyphen in their handle
Full Changelog: https://github.com/av/harbor/compare/v0.3.28...v0.3.29

Onyx
Onyx (formerly Danswer) is an open-source AI platform that combines a Chat UI with powerful RAG capabilities. It's a self-hostable alternative to enterprise knowledge management and AI assistant platforms.
# Start Onyx with Ollama
harbor up ollama onyx --open
Misc
- Fix harbor config update multiplication due to duplicated config in BROWSERUSE service
Full Changelog: https://github.com/av/harbor/compare/v0.3.27...v0.3.28

MindsDB
MindsDB is an abstract data connector for LLMs and integrations to various data providers. It allows to connect, process and query the data with LLMs.
Full Changelog: https://github.com/av/harbor/compare/v0.3.25...v0.3.26

Netdata
Real-time infrastructure monitoring with per-second metrics for systems, containers, and applications.
# Start the service
harbor up netdata --open
Resume Matcher
AI-powered tool for comparing resumes against job descriptions using local LLMs via Ollama.
# Start Resume Matcher with Ollama
harbor up resume-matcher ollama
draw.io
AI-powered diagram creation tool - generate draw.io diagrams from natural language.
# Start with Ollama (default configuration)
harbor up drawio ollama
Misc
- Improved agentic setup
- llamacpp - fix outdated source image - most recent llama.cpp versions are available
Full Changelog: https://github.com/av/harbor/compare/v0.3.24...v0.3.25

Karakeep
Karakeep is an open-source, self-hosted bookmark manager with AI-powered automatic tagging. Save and organize links, notes, images, and PDFs with full-text search powered by Meilisearch.
harbor up karakeep
Full Changelog: https://github.com/av/harbor/compare/v0.3.23...v0.3.24

Presenton
# [Optional] Pre-pull the image
# might take a while as it bundles Ollama + GPU support
harbor pull presenton

# Start the service with web UI
harbor up presenton --open
Misc
- boost - more strategies for klmbr module, fixes for selection
Full Changelog: https://github.com/av/harbor/compare/v0.3.22...v0.3.23

Windmill
Self-hostable alternative to Retool, Pipedream, Superblocks. Low/no-code platform for tools/jobs/workflows. Not specifically oriented towards LLMs, but can work in tandem with other Harbor services.
Misc
- comfyui - improve documentation about workspace sync, re-enable AUTO_UPDATE=true
- llamaswap - switch defautl user to app as per update in the upstream image
- friendlier agentic setup for recent versions of VS Code and Copilot
Full Changelog: https://github.com/av/harbor/compare/v0.3.21...v0.3.22

Unsloth
Unsloth is a Jupyter Lab environment pre-configured with the Unsloth library for fast LLM fine-tuning. It provides 2x faster training with 70% less memory usage compared to standard fine-tuning methods, along with pre-loaded example notebooks for various fine-tuning tasks.
# Pull the pre-built image
# Image is ~7GB which is typical for CUDA-enabled environments
harbor pull unsloth

# Start the service (opens Jupyter Lab in browser)
harbor up unsloth --open
Full Changelog: https://github.com/av/harbor/compare/v0.3.20...v0.3.21

Changelog
- harbor profile use ... - [Profiles] can be imported from a URL: direct, Gist, GitHub, Bitbucket, GitLab, or Pastebin
- Profiles can now be partial - only include the config vars that need an update by the profile
- fabric - using official pre-built image
- browseruse - basic implementation of the service, usable but with some significant caveats
- webui - relaxed healthcheck
- harbor attach <service> - attach to service STDIN/OUT in interactive mode
- harbor up <service> --attach - start service and attach immediately
- harbor profile merge <name> - explicitly merge given profile into the current one
Full Changelog: https://github.com/av/harbor/compare/v0.3.19...v0.3.20

Docling
Docling simplifies document processing, parsing diverse formats — including advanced PDF understanding — and providing seamless integrations with the gen AI ecosystem.
harbor up docling
Misc
- speaches now working again with newer version of the service + automatic model downloads
- harbor config extra options
- harbor pull fixed regression after v0.3.16 not allowing to pull custom services
- harbor open proper handling when system open fails
- harbor ls | ps | ... - correct handling of --no-defaults
- nbs launch example
- bump golang >= 1.24.0 by @genevera in https://github.com/av/harbor/pull/185
- fix doc generation order
New Contributors
- @genevera made their first contribution in https://github.com/av/harbor/pull/185
Full Changelog: https://github.com/av/harbor/compare/v0.3.18...v0.3.19

- boost
- new nbs module
- llm.emit_artifact convenience tweaks
- harbor pull - now pulls Ollama models too: harbor pull gemma3, harbor pull webui
- harbor ollama <command> - now launches Ollama if it's not running yet
- routines migrated to TypeScript
Full Changelog: https://github.com/av/harbor/compare/v0.3.16...v0.3.18

- boost
- full docs revamp: env vars, modules references
- HARBOR_BOOST_MODULES supports "all" option
- discussurl uses Jina Reader instead of raw content
- parllama - moving away from pkgx, fixing config location
- vllm - support for vllm.image config
Full Changelog: https://github.com/av/harbor/compare/v0.3.15...v0.3.16

Airweave

Airweave is a one-stop-shop for integrating external knowledge for use by agents. Current integration is preliminary, awaiting better configurability for embeddings/LLMs in the upstream.
Misc
- boost
- Default temp is 0.35
- polyglot - adjust prompts for smaller LLMs
- chat.advance fixed regression breaking multiple modules relying on this method (rcn and others)
Full Changelog: https://github.com/av/harbor/compare/v0.3.11...v0.3.12

This is a maintenance release with few small bugfixes
- ldr
- fixing env vars for Ollama/SearXNG for newer version
- switching to official prebuilt image
- litellm - config pointers
Full Changelog: https://github.com/av/harbor/compare/v0.3.10...v0.3.11

Modular MAX

MAX is a platform for inference from Modular (creators of Mojo lang). You will need Nvidia GPU to work with this service.
# Pull the image, start the service, tail logs
# ⚠️ Service will download configured model automatically
harbor up modularmax --tail
Misc
- docs
- dev pipeline improvements, enables normalising local MD links in the future (to fix dead local links in the Wiki/repo)
- normalising formatting for the service metadata
- modernised searxng wiki entry
- harbor help / harbor how - CLI help updates
Full Changelog: https://github.com/av/harbor/compare/v0.3.9...v0.3.10

Agent Zero

General-purpose personal assistant with: Web RAG, persistent memory, tools, Browser Use and more.
# --open is optional to open in the browser
harbor up agentzero --open
Misc
- metamcp - fix broken build, point to v0.4.3 (see #156)
- mistralrs - fixing wrong service tag
Full Changelog: https://github.com/av/harbor/compare/v0.3.8...v0.3.9

LocalAI

All-in-one service to download and run LLMs, TTS/STT/Image generation models
harbor up localai
Local Deep Research

Deep Research functionality with local LLMs
harbor up ldr searxng
Harbor Boost - new workflows

- dot - Draft of Thoughts
- grug - Grug make plan with simple word-think. Grug find new path in head-picture that smart-talk no find.
- polyglot - Model generates answers in multiple langugages then synthesises a final solution
- tri - Generates three aspects, then three pitfalls and then three alternative answers before synthesising a final solution
Misc
- boost
- Switching to av/tools, use uv
- llm - stream_completion allows skipping emit to the client
- querying downstream models is cached for 1 minute
- ollama - disabling automatic restarts to align with other Harbor services
- searxng
- aligning configuration with other services - configurable image/version
- fixed healthcheck
- harbor config set searxng.internal_url to point all harbor searxng use to an external SearXNG instance
- librechat - connect to boost
- ollama - modelfile for Gemma 3 QAT w/ tool use in the template
Full Changelog: https://github.com/av/harbor/compare/v0.3.7...v0.3.8

In this release, Harbor gets access to MCP and OpenAPI tools ecosystems with the following new services:
- MetaMCP
- mcpo
# Bring your MCP tools to Open WebUI
harbor up mcpo metamcp
More details in the Harbor Tools guide
Additionally, I'm launching av/tools to simplify containerized tool use in general.
Full Changelog: https://github.com/av/harbor/compare/v0.3.6...v0.3.7

Free and Open Source Machine Translation API, entirely self-hosted.
# [Optional] pre-pull the image
harbor pull libretranslate

# Start the service
harbor up libretranslate
Pulls translation models on the first start - may take a while, monitor with harbor logs libretranslate
Misc
- app
- bundle docs in GH action
- attempt to fix #64
- boost
- basic support for proxying tool calls
- logging revamp
- request ID tracking
- n8n - fix typo in the docs preventing initial workflow import
Full Changelog: https://github.com/av/harbor/compare/v0.3.5...v0.3.6

Harbor App now has dedicated service pages. There are a lot of plans, but we start simple, you'll be able to perform the same actions as in the Home page + see service wiki in the app.
Misc
- harbor dev app helper
- App - bumping Tauri deps
Full Changelog: https://github.com/av/harbor/compare/v0.3.4...v0.3.5

llama-swap
llama-swap is a lightweight, transparent proxy server that provides automatic model swapping to llama.cpp's server.
# [Optional] pre-pull the image
harbor pull llamaswap
# Edit the swap config
open $(harbor home)/llamaswap/config.yaml
# Run the service
harbor up llamaswap
Misc
- boost
- docs revamp
- fixing plain proxy without modules
- tgi - HF cache normalised
- raglite - adding missing traefik config
Full Changelog: https://github.com/av/harbor/compare/v0.3.3...v0.3.4

oterm

the text-based terminal client for Ollama.
harbor up ollama
harbor oterm
RAGLite
⚠️ Unfortunately current integration is not fully compatible with Ollama. See docs for more info and this issue for details.

RAGLite is a Python toolkit for Retrieval-Augmented Generation (RAG) with PostgreSQL or SQLite. Harbor's intergration is centered around the provided chainlit WebUI.
ROCm Ollama
Automatic AMD detection and capability by @cedstrom in https://github.com/av/harbor/pull/143
DND Module for Boost
https://github.com/user-attachments/assets/534e52c4-6e53-4b9e-9ff4-5549c6564299
This module makes LLM pass a skill check before generating a reply to your last message.
App - service names, tooltips

App now displays actual Service names and tooltips with service info. Huge thanks to @cedstrom for the inspiration 🙌
Misc
- boost
- There's now a starter repo for standalone use of Harbor Boost
- Removed incompatible num_ctx from default LLM params
- Fixed streaming for incomplete/merged chunks (seen in Azure/Groq APIs)
- n8n - fixed persistence for custom module installation (broken workspace path)
- docs - service index now generated from app service metadata (now has tags)
- ollama - configure default ctx length with harbor ctx
- webtop - restored Harbor App functionality
New Contributors
- @cedstrom made their first contribution in https://github.com/av/harbor/pull/143 💪
Full Changelog: https://github.com/av/harbor/compare/v0.3.2...v0.3.3

This is a very minor bugfix release
- boost - now correctly handles incomplete chunks from downstream APIs by attempting buffering and then parsing (tested with Groq API)
Full Changelog: https://github.com/av/harbor/compare/v0.3.1...v0.3.2

This is a maintenance release with a few fixes, nothing exciting
- harbor dev docs - fixes relative URLs so that Boost README links now finally work
- README - revamp, supporters
- boost
- fixed mismatch between docs and actual env vars
- r0 - workflow for R1-like reasoning chains for any LLM (including older ones, like Llama 2)
- markov - Open WebUI-only, serves an artifact showing a token graph for the current completion
- docs - numerous tweaks and adjustments
- n8n - fixed missing EOF preventing harbor env n8n from working as expected
- txtai - restored functionality with a monkey patch until this PR is merged
Full Changelog: https://github.com/av/harbor/compare/v0.3.0...v0.3.1

v0.3.0 - Routines, Traefik, Latent Scope
Routines

We now have more than 200 compose files and docker compose itself is being slow merging them (upwards of ~5s even on my powerful dev machine). In order to overcome this we move the core logic that powers the (now legacy) version of the CLI into dedicated routines. v0.3.0 is also a step towards having a native Harbor CLI in the future.
New routines setup is based on the distroless flavor of Deno. In typical Harbor fashion you don't need to install anything, only to pay the disk space tax (~150Mb in this instance). Harbor will cache the dependencies and everything needed after the first cold start. PyPi, Native, and NPM installation paths will continue to function in the same way as previously. Deno was chosen over Bun, Node.js, Python because it brings the most value within these 150Mb with a path towards native binaries in the future and much more. Additionally, Harbor was already using Deno for some lightweight automation. I did experiments with Rust and Go and despite the better end binaries - the cost of creating and maintaining a CLI there is much higher, so choosing Deno means more time to update and improve the project.
You can return to the legacy behavior by setting a config option:
harbor config set legacy_cli true
Traefik
Harbor now includes traefik as its reverse HTTP Proxy. Automatic configuration is now limited to a local deployment, however it can be reconfigured as needed manually - see the service wiki for more details.
Latent Scope

One of the tools that leave you with a "woah". Allows exploring the given dataset representation in latent space.
Misc
- perplexideez - fixed usage without access to modern compose
- release script seeds programmatic files
- seed-cdi - EOF fixes
- seed-traefik - dev script to create .x.traefik. cross-files for related services
- boost - revised r0 module
- qrgen - fixed build (affects harbor qr, harbor tunnel)

New Contributors
- @heronsouzamarques made their first contribution in https://github.com/av/harbor/pull/134 💪
- @Tien-Cheng helped restoring the qrgen functionality in https://github.com/av/harbor/pull/137
Full Changelog: https://github.com/av/harbor/compare/v0.2.28...v0.3.0

This is a mostly maintenance release with a small new Frontend and an exciting new feature for Harbor Boost.
Mikupad

LLM frontend in a single HTML file
Misc
- boost
- support for interactive artifacts (specifically for Open WebUI),
- dedicated README
- multiple new experimental modules (undocumented, see source)
- relaxing CORS policy
- cosmetic fixes to dev scripts
- librechat - fixing after MeiliSearch update to v1.12.3
- promptfoo
- access to Harbor's configured API keys
- example eval based on Misguided Attention (unfinished)
- webui - now sees local time correctly on supported systems
- harbor logs, harbor down and other commands relying on "*" - fixed incorrect detection of CDI capability
New Contributors
- @ColumbusAI made their first contribution in https://github.com/av/harbor/pull/131 🙌
- @kianmeng made their first contribution in https://github.com/av/harbor/pull/124 💪
Full Changelog: https://github.com/av/harbor/compare/v0.2.27...v0.2.28

Three new services are here to add even more value to your local LLM setup!
Morphic

An AI-powered search engine with a generative UI.
SQL Chat

SQL Chat is a chat-based SQL client, which uses natural language to communicate with the database to implement operations such as query, modification, addition, and deletion of the database.
gptme

Terminal assistant, with tools so it can: use the shell, run code, edit files, and much more;
Speaches - now supports Kokoro v1
Harbor patched its installation of speaches to support Kokoro v1 models (ahead of official support from the project itself)
Misc
- CDI - Harbor detects and enables CDI Nvidia driver on compatible systems (kudos to @FrantaNautilus)
- harbor dev - alias to run dev-related scripts from .scripts
- harbor dev scaffold - unwanted prefix newline
- boost - extras for chat/chat node APIs for custom modules
- cex - experiment on automatic context expansion by paraphrasing
- stcl - continued experiments on "side" reasoning
- Plenty of clarifications and extra-examples for Ollama/WebUI docs and more
- Experimental requirements.sh to install Harbor's dependencies automatically on Linux (undocumented, untested)
New Contributors
- @FrantaNautilus made their first contribution in https://github.com/av/harbor/pull/119 🎉
Full Changelog: https://github.com/av/harbor/compare/v0.2.26...v0.2.27

This is a maintenance release with a bugfixes for specific services and general cross-platform compatibility
- cmdh
- stop using pkgx, patch in structured Outputs from Ollama and larger default context (should be usable with llama3.1 8b)
- harbor how will actively check if Ollama has the model configured for cmdh and will ask to pull otherwise
- ollama - extra docs on cache location, extrenal instance and troubleshooting
- boost
- continuing experiments with stcl workflow
- @bjj fixed handling of Docker Desktop versions for harbor doctor and capability detection
- fixed incorrect application of capabilities when using a wildcard * for service match
- This was the reason some services were failing to launch on Mac OS
- aider - fixes to support non-root user in newer versions
- openhands - fixes to support custom registry, local state volume in newer versions
- librechat - fixed entrypoint lacking exec permissions
- shared - Node.js config merger no longer requires lodash in the base container
- Fixes chatui being unable to start in recent versions
- bolt - switch to official bolt.diy image, notes on Mac OS compatibility in the docs
New Contributors
- @bjj made their first contribution in #112 🎉
Full Changelog: https://github.com/av/harbor/compare/v0.2.25...v0.2.26