convez.ai
MCP Server
OwLLM

owllm

OwLLM/owllm

Ver no GitHub
BomPluginScore · de 0 a 100

Local-first AI workstation. Run open-weight models, fine-tune, orchestrate multi-agent teams. No cloud required.

Instalação

$npx convez-ai OwLLM/owllm

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
87%
Último commit
há 1 dia
Versão
v1.0.93
Issues abertas
1
Contribuidores
2

README

A tradução para o português deste README ainda está na fila.

⬇ Download OwLLM for Windows

One file · ~30 MB · No admin required · Windows 10 / 11 x64

installer portable


Your team of AI agents. Build them. Own them. Run them anywhere.

OwLLM is an open platform to build, deploy, and run custom AI agent teams — on your hardware, your VPS, or in a VM, 24/7. Bring your own models: local, cloud, or both. Fine-tune. Quantize. Abliterate. Red-team. Automate.

[!IMPORTANT] OwLLM Desktop currently ships for Windows 10/11 (x64) only. macOS (Apple Silicon + Intel) and Linux (x86_64) builds are coming via the already-configured cross-platform CI. Watch the repo for release notifications.


What makes OwLLM different

Most AI tools give you a chatbox. OwLLM gives you a workforce.

You compose teams of specialised agents — an orchestrator that plans, a coder that writes, a critic that reviews, a researcher that fact-checks — and they collaborate on real tasks in parallel. Each team is a graph of roles + prompts you define. The 18 teams shipped in this repo are starter samples, not the menu.

What OwLLM gives you that others don't
🧩 Build your own teamsCompose agents from 8 base roles + custom prompts. Visual graph builder. Hot-updates through this repo — push a team JSON, it lands on every installed app.
☁️ Cloud OR local — same teamsNo 4090? Plug in Claude / GPT / Gemini / Kimi keys, teams work identically. Have a GPU? Run open-weight models locally and stop paying per token. Mix both in the same conversation.
🎓 Fine-tune any modelFull LoRA + Unsloth + TRL pipeline. Drop a JSONL, watch loss curves, save adapters. Works on consumer GPUs (8 GB+).
🔬 Abliterate for safety researchOrthogonalise weights against refusal directions. Generate adversarial datasets. Train better safety classifiers. The honest tools the field actually needs.
🛠 GGUF + quantization built-inConvert HF safetensors → GGUF, quantize Q4/Q5/Q6/Q8/F16. Ship custom models anyone with llama.cpp can run.
🛡 Red-team capableCompose adversarial agent teams whose job is to find vulnerabilities — in models, code, apps. Pair with fine-tuning to train defenders.
🔒 Sealed to ONE folder (Win/Mac/Linux)Flip it on and every tool your agents run — shell, file writes, edits, search, and the cloud CLIs (Claude/Codex/Gemini/Kimi) — runs inside a real Linux sandbox (WSL2 on Windows, Lima VM on macOS, bubblewrap on Linux (Mac/Linux beta)). Agents are sealed to ONLY the project folder: they work on your real Windows folder — no copy — but cannot see the rest of your C: drive, other projects, or your home / SSH keys. A model that runs rm -rf simply has nothing else to reach. Isolated by default; the seal and toolchain auto-install; provider logins (CLIs and API keys) auto-sync in so cloud agents just work. Connect GitHub to clone/push private repos from inside. See the 🔒 Sealed to one folder section below for the diagram.
🔌 MCP-first toolingPlug in any Model Context Protocol server (filesystem, git, browser, Postgres, GitHub…). Keyless DuckDuckGo web search is auto-installed on first run — no API key, no card. Engine-agnostic: any search MCP you add is used automatically. Curated packs per team.
🏠 Run anywhereDesktop today. Headless on a $5/mo VPS, 24/7 — on the roadmap. Containerised / VM — on the roadmap. Your agents, your hardware, your terms.

🔒 Sealed to one folder — nothing else

Give an AI agent a shell and it can read your SSH keys, wander into other repos, or rm -rf the wrong thing. OwLLM puts a hard wall around that. Turn isolation on (it's the default) and every command runs inside a Linux sandbox that can see only the project folder you're working on — your real Windows folder, mounted live with no copy — and nothing else of your machine.

flowchart LR
    A["🤖 Agent<br/>shell · edits · cloud CLIs<br/>(Claude · Codex · Gemini)"]
    A ==>|"sealed by bubblewrap inside WSL2"| BOX

    subgraph BOX["🔒 The sandbox sees ONLY this"]
        direction TB
        P["📁 Your project folder<br/>your real Windows files — no copy"]
        K["🔑 The agent's own logins<br/>auto-synced — cloud agents just work"]
    end

    BOX -.->|"🚫 blocked"| C1["🗂️ Rest of your C: drive"]
    BOX -.->|"🚫 blocked"| C2["📦 Your other projects"]
    BOX -.->|"🚫 blocked"| C3["🔐 Home · SSH keys · browser"]

    classDef safe fill:#0e2a1e,stroke:#3ec58a,stroke-width:2px,color:#d6ffe9;
    classDef block fill:#2a1414,stroke:#d86a6a,stroke-width:1px,color:#ffd9d9;
    classDef agent fill:#0a2230,stroke:#3ec5d8,stroke-width:2px,color:#dff6ff;
    class A agent;
    class P,K safe;
    class C1,C2,C3 block;
    style BOX fill:#0c1f17,stroke:#3ec58a,stroke-width:2px,color:#d6ffe9;
  • Real folder, no copy — work on C:\code\my-repo directly; changes are immediate, nothing is duplicated, no disk doubling.
  • Keeps working after you seal it — the agent's Claude/Codex/Gemini logins are bound in, so the team doesn't suddenly ask you to log in again.
  • One toggle, honest feedback — press 🔍 Verify and the app runs a probe through the agent's own shell and tells you plainly: runs in WSL? sealed to only this folder? — and installs the seal if it's missing.

💾 …and it cleans up after itself

The sandbox lives in WSL, whose virtual disk only ever grows. A Sandbox disk card on the Home page keeps it honest:

ActionWhat it doesCost
See usageWSL disk size + reclaimable caches (uv / npm / pip) + project-copy size, at a glance—
🧹 Clear cachesDrops regenerable build caches — never your projects, logins, or models. Often frees many GB in secondsinstant, no restart
💿 Reclaim diskPhysically shrinks the WSL disk file and reports exactly how much came backrestarts WSL · admin prompt

Deleting a project also frees its sandbox copy automatically — your own folders are never touched.

What teams can do

OwLLM ships starter teams in nine categories. All of them are forkable and remixable — they're templates, not the menu. The real product is the team builder.

CategoryWhat teams here doStarter samples
🛠 CodeArchitect → code → critic → refactor; bug hunting; reviewscode_artisan, dev_squad, code_reviewer, bug_hunter
🔬 ResearchMulti-source synthesis with real citations, fact-checkingresearch_lab, learning_tutor
📊 DataSQL → notebook → viz → narrativedata_analyst
🎨 DesignProduct → UX → tech → critiqueproduct_studio
✍️ WritingOutline → draft → edit → SEO → publishwriters_room, social_desk
🤝 OpsTriage → respond → schedule → digestsecretary, concierge, customer_support
💼 PersonalCalendar, finance, health, home automationfinance, health_coach, smart_home
🌐 SocialOutreach, support, community managementsales_outreach, n8n_workflow_builder
🛡 Safety / Red-teamAdversarial dataset generation, jailbreak research, refusal probing(build your own — see data/teams/SCHEMA.md)
🎮 GamifyAgent-vs-agent, achievements, arena(in progress — Q4 2026)

Browse the 18 starter teams → · Build your own →

Build your own team — 5-minute walkthrough

  1. Open Studio in the desktop app
  2. Drop in agents: orchestrator + 1..N specialists (coder, critic, researcher, brainstormer, devops, documentation, operator, …)
  3. Wire the dispatch graph (orchestrator → coder → critic → back to orchestrator)
  4. Write each agent's system prompt
  5. Save → team appears in your picker
  6. Publish to the community via PR against data/teams/ — your team becomes one-click installable for every other user

Power tools nobody else ships

Fine-tune any open-weight model

LoRA pipeline with Unsloth, TRL, PEFT, bitsandbytes. Llama / Qwen / Mistral / Gemma — anything on HuggingFace. Live loss curves, graceful Stop preserves checkpoints, resume-from-checkpoint and resume-adapter both supported. Runs on a 12 GB GPU.

Abliterate (refusal removal for safety research)

Orthogonalise weight matrices against refusal directions (the Labonne / Arditi technique, packaged). Use cases:

  • AI safety labs training refusal classifiers need cleanly-uncensored teacher models
  • Red teams need models that don't sandbag jailbreak tests
  • Academic research on alignment failure modes

The corpus prep + abliteration script ship together.

GGUF creation + quantization

Convert HF safetensors → GGUF, quantize to Q4_K_M / Q5_K_M / Q6_K / Q8_0 / F16. The same pipeline that gives you tiny, fast custom models others can run on llama.cpp / Ollama / LM Studio.

Adversarial dataset generation

Build a team whose role is to PROBE another model. Output: a labelled dataset of jailbreak attempts, refusal patterns, edge cases. Sells to AI safety labs. Trains your own filters.

Cloud or local — same teams, your choice

You don't need a 4090. Many users will never have one.

  • Cloud-only: Plug in Claude / GPT / Gemini / Kimi API keys. Teams work identically. ~30 MB install, runs on any laptop.
  • Local + cloud mix: Have a 3060? Run Llama for the bulk, hand off to Claude for the hard parts in the same conversation. Save 90% on tokens.
  • Local-only: Have a 4090? Never touch a cloud API. Privacy by default. Stop paying per token forever.

Same teams. Same agent definitions. Same UI. The model layer is just plumbing.

Run anywhere

ModeStatusUse case
Desktop (Windows)✅ shippedDaily-driver AI workstation on your laptop
Desktop (macOS / Linux)🔜 Q3 2026Mac / Ubuntu users
Headless on VPS (24/7)🔜 Q4 2026Run your custom teams on a $5/mo box. Reach them via Telegram, web, API. Always-on agentic services.
Containerised / VM🔜 Q4 2026Drop OwLLM into your existing infra.

The team definitions, role prompts, MCP configs, and model selections are all portable across deployment modes — build a team once, run it anywhere.

Install (Windows only — for now)

  1. Download OwLLM.Desktop.Setup.exe (~30 MB — one file, that's it)
  2. Run OwLLM-Desktop-Setup-x64.exe. Windows SmartScreen may flag it the first time (the binary isn't EV-signed yet) — click "More info" → "Run anyway".
  3. On first launch, a hardware-aware wizard opens. It detects your hardware and offers the modules that fit:
    • Local Inference (~33 MB CPU / ~32 MB Vulkan / ~285 MB CUDA) — only needed if you want local models
    • Audio / Speech-to-Text (~148 MB) — for voice messages, mic input
    • Fine-tuning (~12 GB) — only if you'll train models
    • MCP toolchain (~260 MB) — only if you want browser / git / postgres MCP servers

Cloud-only? Skip the wizard entirely and just enter your API keys in Settings. The shell alone is enough for cloud-model chat + agent orchestration.

How updates work

Three independent update streams — small, fast, no full reinstalls:

  • Shell auto-updates via Tauri's signed updater
  • Modules (llama backend, fine-tune env, audio, MCP) check + swap per-launch
  • Data layer (team templates, role prompts, model profiles, MCP recommendations) hot-pulls from data/ in this repo on launch. A new team you contribute today reaches every installed app within minutes — no rebuild.

That's why the data/ tree is open and community-driven even though the app binaries are closed-source.

Roadmap

  • Multi-agent dispatch with worktree isolation
  • Modular installer + hardware-aware wizard
  • MCP-first tool architecture
  • Fine-tuning + abliteration pipeline
  • GGUF / quantization pipeline
  • Telegram bridge
  • WSL tool isolation — agents run their tools inside Ubuntu, off your Windows drive
  • Folder-sealed isolation — agents see only the project folder (bubblewrap inside WSL), on your real Windows folder, no copy
  • Sandbox disk management — usage view, one-click cache clear, disk reclaim; auto-cleanup on project delete
  • Cloud CLIs inside the sandbox — Claude/Codex/Gemini/Kimi run isolated too
  • Connect GitHub — isolated agents clone private repos + push from inside the sandbox
  • Auto login-sync — codex/claude/gemini/kimi + every API key mirrored into the sandbox
  • Convert projects isolated↔not from the header; Accounts tests host + sandbox
  • [~] Mac/Linux isolation (beta) — Lima VM (macOS) + bubblewrap (Linux), same model as WSL
  • Visual team builder — Q3 2026
  • macOS + Linux desktop — Q3 2026
  • 24/7 headless / VPS mode — Q4 2026
  • Container / VM deployment — Q4 2026
  • Gamification (agent-vs-agent arena, achievements) — Q4 2026 (in progress)
  • WhatsApp bridge — Q4 2026
  • Vision models (LLaVA / Pixtral) — Q4 2026
  • Voice output (TTS) — Q1 2027
  • Public team marketplace — Q1 2027

Track active work in Discussions → Roadmap.

Who's this for

  • Indie devs & founders — your AI workforce, not a SaaS subscription
  • AI safety researchers — abliteration, red-team teams, adversarial dataset gen
  • Model creators — fine-tune, quantize, ship GGUFs
  • Automation builders — replace n8n / Zapier with agents that understand meaning
  • Privacy-bound teams — legal, medical, defence, regulated industries
  • Agencies — run custom client agent teams 24/7 (when VPS mode lands)
  • Power users — anyone tired of generic chatboxes

Community

License

Repository contents (agent teams, role definitions, registry, schemas, docs): MIT — fork freely, share team packs, build on it.

Application binaries via Releases: see EULA.md. Source for the application itself is not currently public.

Acknowledgements

Standing on the shoulders of: llama.cpp, whisper.cpp, Tauri, Unsloth, Model Context Protocol, and the open-weight model creators (Meta, Alibaba, Mistral, Google, DeepSeek, Anthropic for their safety research).

If you build something cool with OwLLM, share it in Discussions → Show & Tell. Stars are how this category proves itself worth investing in.

Ver o original no GitHub

Bom para

Relacionados