archestack 发布 best-of-selfhosted-ai：自托管 AI 项目排行榜

Self-hosted AI, ranked. 🏆

What runs on your own box, what it needs, where it breaks. No fluff, and paid never moves a rank.

Tip

🤖 Bots find and fact-check every project  ·  ✍️ Every entry has strengths and weaknesses  ·  ▶️ Demo links so you can try before you deploy

Looking for starters and templates to build your own AI app? 👉 Best of AI Starters

🔥 Top 10 right now

_{Ordered by score (how we rank): adoption, freshness, how easy it is to run, and a few more signals as the bots measure them. Expect the board to shuffle. 🔀}

⚡ Pick by vibe

• 💬 ChatGPT, but at home → Chat UIs
• 🧠 Run models on my own GPU → Model serving
• 🤖 An assistant living in my Telegram or WhatsApp → Assistants
• 📚 Chat with my own documents → RAG and knowledge
• 🧩 Build agents and workflows without coding everything → Agent platforms
• 🔀 One API key for every provider → Gateways
• 💻 Copilot, self-hosted → Coding
• 🎙️ Voice in, voice out, offline → Voice
• 🎨 Generate images and video → Image and video
• 🔎 Search that doesn't phone home → Search
• 📈 See what my LLM app is actually doing → Observability
• 🛡️ Let agents run code without wrecking my server → Sandboxes

🗂️ All categories

• 🤖 Assistants · 11
• 💬 Chat UIs · 13
• 🧩 Agent platforms · 12
• 📚 RAG and knowledge · 15
• 🧠 Model serving · 18
• 🔀 Gateways · 13
• 🗂️ Memory · 11
• 🎙️ Voice · 11
• 🎨 Image and video · 8
• 💻 Coding · 9
• 🔎 Search · 8
• 📈 Observability · 13
• 🧮 Vector databases · 9
• 🛡️ Sandboxes · 8

_{Legend: 🥇 80+ · 🥈 65–79 · 🥉 55–64 · number = score out of 100 (how we rank) · ⭐ GitHub stars · 📝 review (strengths, weaknesses, specs) · ▶️ live demo · 📖 docs · 🌐 website (↗ = leaves GitHub; GitHub ignores target=_blank, so Cmd/Ctrl-click for a new tab) · 🐳 Docker image or compose · 🎮 GPU required · ✨ GPU optional}

🤖 Assistants

Personal AI assistants you run yourself and talk to through chat apps, with memory and the ability to act. _{11 projects · 📝 all reviews}

💡 How to choose

• Check which channels it supports today (WhatsApp, Telegram, Slack, Discord, email) and whether each needs a paid API.
• Look at how memory is stored and whether you can inspect or wipe it.
• Actions need credentials; prefer assistants that scope them per tool.

↑ categories

💬 Chat UIs

Web front-ends for local or API models, usually with user accounts, chat history and file upload. _{13 projects · 📝 all reviews}

💡 How to choose

• Confirm it talks to your backend (Ollama, OpenAI-compatible, Anthropic) without a plugin.
• Multi-user auth, RBAC and SSO are where free and paid editions differ most.
• Check what the default compose pulls in (database, vector store) before sizing the host.

↑ categories

🧩 Agent platforms

Visual or code-first builders for agents and workflows, with orchestration, tools and deployment. _{12 projects · 📝 all reviews}

💡 How to choose

• Decide between a visual builder (faster to start) and code-first (easier to test and version).
• Check how workflows are exported; vendor-specific JSON makes migration costly.
• Look at the license for the server part; several are AGPL or source-available.

↑ categories

📚 RAG and knowledge

Document Q&A, knowledge bases and enterprise search over your own files and data. _{15 projects · 📝 all reviews}

💡 How to choose

• Match the ingestion formats you need (PDF with tables, Office, web, Confluence) before anything else.
• Check which embedding models and vector stores are supported and whether they can run offline.
• Look for citations in answers; without them, RAG output is hard to trust.

↑ categories

🧠 Model serving

Inference engines and model servers that expose local models over an API. _{18 projects · 📝 all reviews}

💡 How to choose

• Pick by hardware first (CPU, NVIDIA, AMD, Apple Silicon) and by model format (GGUF, safetensors).
• Throughput engines (continuous batching, paged attention) need GPUs; single-user servers do not.
• An OpenAI-compatible API keeps the rest of your stack portable.

↑ categories

🔀 Gateways

LLM gateways and proxies for routing, caching, rate limits and cost control across providers. _{13 projects · 📝 all reviews}

💡 How to choose

• List the providers you call today; check the gateway's native support rather than generic passthrough.
• Key management, per-team budgets and audit logs separate gateways from simple proxies.
• Check latency overhead and whether streaming is passed through unchanged.

↑ categories

🗂️ Memory

Long-term memory engines that store and retrieve facts for agents across sessions. _{11 projects · 📝 all reviews}

💡 How to choose

• Check what gets stored (raw messages, extracted facts, graphs) and how it is retrieved.
• Multi-tenant isolation matters if several users or agents share one store.
• Look at the backing store (Postgres, a vector DB, a graph DB) and whether you already run it.

↑ categories

🎙️ Voice

Speech-to-text, text-to-speech, voice agents and meeting tools that run locally. _{11 projects · 📝 all reviews}

💡 How to choose

• Latency decides usability for voice agents; check the end-to-end numbers the project publishes.
• Most quality models need a GPU; CPU-only setups are slower and limited to smaller models.
• Check language coverage and whether models are downloaded at build or at first run.

↑ categories

🎨 Image and video

Generation UIs and pipelines for images and video, usually around diffusion models. _{8 projects · 📝 all reviews}

💡 How to choose

• A GPU with enough VRAM is the hard requirement; check the minimum the project states.
• Node-based UIs are flexible but have a learning curve; form-based UIs are quicker to use.
• Check the model licensing separately from the app licensing.

↑ categories

💻 Coding

Self-hosted coding assistants and agents, from editor completion to autonomous task runners. _{9 projects · 📝 all reviews}

💡 How to choose

• Separate editor completion (needs low latency, small models) from agents that run tasks (need strong models).
• Check sandboxing for agents that execute code or shell commands.
• Look at which editors and model backends are supported without a cloud account.

↑ categories

🔎 Search

Private search engines and AI answer engines that keep queries on your host. _{8 projects · 📝 all reviews}

💡 How to choose

• Metasearch engines need upstream providers; check rate limits and whether results are cached.
• Answer engines call an LLM per query; budget tokens before exposing them to a team.
• Check how results are attributed; answer engines without citations are hard to verify.

↑ categories

📈 Observability

Tracing, evaluation and prompt management for LLM applications. _{13 projects · 📝 all reviews}

💡 How to choose

• Confirm SDK support for your framework (OpenTelemetry, LangChain, OpenAI SDK) and your language.
• Evaluation features vary widely; check whether evals run on your data without a cloud account.
• Retention and storage backend decide the host size; traces grow fast.

↑ categories

🧮 Vector databases

Vector stores and hybrid search engines for embeddings. _{9 projects · 📝 all reviews}

💡 How to choose

• If you already run Postgres, check pgvector-based options before adding a new database.
• Hybrid search (keyword + vector) and filtering are where engines differ most.
• Check memory use per million vectors; it decides the host size.

↑ categories

🛡️ Sandboxes

Isolated runtimes where agents execute code, browse or use tools safely. _{8 projects · 📝 all reviews}

💡 How to choose

• Isolation level (container, microVM, gVisor) sets the risk you accept when agents run code.
• Check startup time per sandbox; slow starts limit agent loops.
• Look at what is persisted between runs and how secrets are injected.

↑ categories

🏅 How we rank

Every project gets a score out of 100 from the signals below. A signal the bots cannot measure yet is left out and its weight goes to the others, so nobody loses points for things we have not checked. 🥇 80+ · 🥈 65–79 · 🥉 55–64.

_{Penalties: no license −10 · data must leave your host −10 · star-growth anomaly −15 · unpatched security advisory −20. Archived projects are removed. Paid placements never touch the score.}

🧠 How this works

• 🔎 Found by bots from curated lists, app stores and template galleries, then fact-checked on GitHub: stars, last commit, license, Docker files.
• ✍️ Written from the README, never copied from other lists: what it does, what it needs, strengths and weaknesses as claims you can check. Specs say unknown rather than guess.
• 🔄 Kept fresh: entries are rewritten when the README or the latest release changes; projects quiet for 12 months get marked stale, archived ones drop off.
• 🚫 No pay-to-rank. Sponsors and affiliate links, if any, are labeled and never touch the order.

📬 Your turn

➕ Add a project · 🛠️ Fix wrong data · 🚪 Opt out (done within 24 hours). Everything here is generated, so use the forms instead of editing files.

📚 Where candidates come from

These lists, app stores and galleries (facts and links only, no text copied), plus community submissions: awesome-selfhosted · awesome-local-llm · compose-examples · awesome-homelab · self-hosting-guide · awesome-ai-agent-platforms · awesome-openclaw · umbrel-apps · runtipi-appstore · casaos-appstore · dokploy-templates · awesome-opensource-ai · awesome-llm-services · awesome-llmops · awesome-private-ai · awesome-local-llms · awesome-llm-webapps.

📄 License

Data (data/, this README, reviews/) is CC BY 4.0; see LICENSE-DATA. Code is MIT; see LICENSE. Project names and descriptions belong to their owners.

_{Made with 🤖 + ☕ by Archestack · see also Best of AI Starters}