metis

Local-First / Self-Hosted AI Platforms (non-IDE)

Verified 2026-05-09. GitHub repo metadata verified live via gh api; feature claims from each project’s public docs.

This stream excludes coding-IDE competitors (Cursor, Cline, Aider, etc.) — those are covered in 01-coding-agents.md. This is the broader self-hosted / local stack.

Per-product matrix

Product	Stars	Surface	Local models	Remote providers	Agent loop?	Tools/MCP	Memory	Multi-provider routing	License	URL
Open WebUI	136,297	Web (self-hosted)	Ollama-native, llama.cpp via OAI-compatible	OpenAI-compatible APIs	Limited — “tools” + “functions” + pipelines, mostly single-turn function calls; not a true autonomous loop	Custom Python “tools/functions”; MCP via OpenAPI bridge (“MCPO”)	Per-conversation; basic “Memories” feature	Manual model switch	OSS (BSD-3 w/ branding clause)	openwebui.com
LibreChat	36,785	Web (self-hosted)	Ollama, OAI-compatible	OpenAI, Anthropic, Bedrock, Vertex, Groq, Mistral, OpenRouter	Yes — “Agents” with multi-step tool calls	OpenAPI Actions, Code Interpreter (paid), MCP servers	Per-conversation; recent “Memory” feature	Manual; per-agent endpoint	OSS (MIT)	librechat.ai
Jan	42,443	Desktop (Tauri-like) + local server	llama.cpp/Cortex bundled	OpenAI, Anthropic, Groq, etc. via extensions	Partial — assistants + tool extensions; not strong autonomous loop	Extensions/MCP added in 2025	Per-thread	Manual	OSS (AGPL)	jan.ai
AnythingLLM	59,776	Desktop + Web (self-hosted)	Ollama, LM Studio, LocalAI, llama.cpp	OpenAI, Anthropic, Bedrock, Gemini	Yes — “@agent” mode w/ tool calls	Built-in skills (web search, scrape, SQL), custom skills, MCP servers	Per-workspace RAG + thread	Manual per-workspace	OSS (MIT, w/ commercial add-ons)	anythingllm.com
LM Studio	n/a (closed)	Desktop + local OAI-compatible server	MLX, llama.cpp native	None native (it IS the provider)	Recent: tool-use + MCP host added (2024–25)	MCP client support	Per-chat	n/a	Proprietary, free	lmstudio.ai
Big-AGI	6,967	Web	Ollama, LocalAI	Most major APIs	Limited — “Beam” multi-model, persona-driven; not deep tool loops	Some browser/code tools	Per-chat	Yes (Beam)	OSS (MIT)	big-agi.com
Letta (ex-MemGPT)	22,570	Server + SDK + ADE web	via Ollama/vLLM	OpenAI, Anthropic, etc.	Yes — stateful agent loops are the core thesis	Tool registration, MCP	Hierarchical bounded memory (core/archival/recall)	Yes, per-agent	OSS (Apache 2.0) + cloud	letta.com
Open Interpreter	63,441	CLI + local server	Ollama, llama.cpp via LiteLLM	All via LiteLLM	Yes — original “do-anything-on-your-computer” agent loop	Shell/Python execution; MCP recently	Conversation only	Yes via LiteLLM	OSS (AGPL); last push 2025-05-27 — stalled ~12 months	openinterpreter.com
gptme	4,294	CLI (+ small web UI)	Ollama, llama.cpp via OpenAI-compat	OpenAI, Anthropic, Gemini, etc.	Yes — terminal agent loop (shell, files, browser, Python, patch)	Built-in tools; MCP support added	Local conversation logs (markdown)	Yes	OSS (MIT)	gptme.org
LocalAI	46,160	Server (OAI-compatible)	Massive (LLM/vision/voice/image)	n/a (it’s a backend)	No — inference engine, not an agent	Just an API	n/a	n/a	OSS (MIT)	localai.io
GPT4All	77,365	Desktop	Bundled GGUF	Limited	No agent loop; chat + LocalDocs RAG	None real	Per-chat	No	OSS (MIT)	gpt4all.io
PrivateGPT	57,211	Server + minimal UI	llama.cpp/Ollama	Optional	No — RAG over docs (last push 2025-05-27 — stagnating)	None	Vector index	No	OSS (Apache 2.0)	privategpt.dev
Ollama	171,064	CLI + local server	Native (GGUF)	n/a	No — model runtime only; tool-call API surface exposed for clients	n/a	n/a	n/a	OSS (MIT)	ollama.com

“Self-hosted but not really” call-outs

• Open Interpreter — last pushed 2025-05-27; effectively stalled ~12 months. Treat as historical.
• Letta has a self-hostable OSS core, but the polished ADE/agent runtime is pushed toward Letta Cloud.
• PrivateGPT repo also pushed 2025-05-27 — also stagnating.
• Open WebUI “tools” are mostly single function-call passes; not an autonomous agent harness despite marketing.
• Big-AGI “deploy on-prem or cloud” — local model support is one Ollama URL field, no orchestration.

What actually overlaps with metis’s positioning?

metis = local-first dev assistant with skills + bounded memory + multi-provider routing + canonical message format. Filtering:

• Genuine overlap (agent loop + tools + persistent state on user’s box): Letta, gptme, AnythingLLM, LibreChat (Agents), Open Interpreter (historical).
• Just chat UIs for local models (not real overlap): Open WebUI, Jan, GPT4All, Big-AGI, LM Studio, PrivateGPT, LocalAI, Ollama.

Closest two

1. gptme — closest spiritually. CLI-first local agent, markdown-logged sessions, shell/file/patch tools, multi-provider via litellm-style abstraction, MCP support, OSS. Differs from metis on: no skills standard, no bounded-memory model, no multi-surface client/server split.
2. Letta — closest on the memory axis. Hierarchical bounded memory (core/archival/recall) is exactly the design pattern metis needs to study; provider-agnostic; tool/MCP capable. Differs: server/cloud-leaning, not dev-tool focused, no skills-as-markdown concept.

(AnythingLLM is a strong third — agent skills, MCP, multi-provider, desktop+server — but it’s RAG-workspace shaped, not dev-assistant shaped.)

Things metis can learn

• Letta’s memory hierarchy (core block / archival / recall + self-edit tools) — directly applicable to metis’s bounded-memory design; their “memory blocks as editable context” pattern is shipping.
• gptme’s markdown-as-source-of-truth session log — sessions are replayable plain files; aligns with metis’s markdown+git sync model.
• AnythingLLM’s “skills” pane — ergonomic UX for enabling/disabling agent capabilities per workspace; useful pattern for agentskills.io toggles in the TUI/dashboard.
• LibreChat’s provider abstraction + per-agent endpoint binding — clean canonical-message bridging across OpenAI/Anthropic/Bedrock/Vertex; good prior art for metis’s canonical message format and routing.
• Open WebUI’s “Pipelines” filter chain — pre/post-LLM hooks (PII redaction, logging) — a pattern metis could expose for trace/skill middleware without baking it into the loop.
• MCPO bridge (Open WebUI) — exposing MCP servers as OpenAPI for non-MCP clients; useful since metis Phase 3 is MCP-centric and clients differ.
• Jan’s local-first desktop branding — the privacy/offline narrative is winning shelf space; metis can borrow positioning language.

Strategic takeaway

The non-IDE local-AI market is dominated by chat UIs, not dev agents. The dev-agent niche on the local-first side is thin (gptme + the dormant Open Interpreter), so metis has real room. Competitive pressure on metis comes from coding-IDE tools (separate stream), not from this category — except on memory architecture, where Letta is meaningfully ahead and worth tracking.

This site is open source. Improve this page.