Who guaardvark is for#
Developers building local AI pipelines with RAG and agents
Guaardvark provides hybrid BM25 and vector RAG over local documents, a coding agent with Monaco editor and AST-aware indexing, and a swarm of up to 20 parallel agents in isolated git worktrees. All inference stays on the machine, so proprietary codebases never leave the local environment.
Skip if:
If your codebase lives in a cloud CI system and your team is comfortable with hosted LLM APIs, a lighter tool like LangSmith gives you agent observability without the infrastructure overhead.
Creative teams producing video and media with local models
The platform runs Wan 2.2, CogVideoX, and LTX for video; Stable Diffusion with bulk CSV pipelines for images; and ACE-Step for music generation. A 5-role Film Crew agent (script, casting, shots, keyframes, edit) can turn a logline into a finished video clip on a single GPU.
Skip if:
If you need the breadth of ComfyUI's node graph for custom diffusion workflows, use ComfyUI directly. Guaardvark can hand off to it with one click, but its built-in media pipeline covers fewer custom workflow nodes.
Privacy-focused teams needing offline-first AI
Flight Mode gives fully offline agent runs, swarm tasks, and RAG over documents, tested end-to-end without internet. Combined with the MCP server's strong default-deny policy and no telemetry by default, Guaardvark suits teams handling sensitive data who cannot use cloud AI APIs.
Skip if:
If you need enterprise compliance certifications (SOC 2, HIPAA BAA), Guaardvark's self-managed nature means you run the compliance audit yourself. Azure AI Studio and Google Vertex carry their own certifications.
Small teams automating social and web outreach locally
The supervised social outreach module drafts comments and replies for Reddit, Discord, and forums, queuing them for explicit approval before posting. A kill switch, cadence limits, per-platform persona, and a JSONL audit trail are built in. All content generation runs on local models.
Skip if:
If your outreach volume requires parallel posting at scale, the supervised queue model may be a bottleneck. The module is designed for thoughtful, reviewed outreach rather than bulk automation.
The problem it solves#
Building with AI agents today means spreading work across a dozen separate services: a managed agent framework, a hosted vector database, a cloud LLM API, a separate video generation endpoint, and a voice assistant with its own subscription. Each service adds a billing account, a data-sharing agreement, and a content filter that can block legitimate use cases.
The deeper pain is data sovereignty. Every document you send to a RAG pipeline, every codebase your coding agent reads, and every image your creative pipeline generates passes through a third party's servers. For teams handling proprietary code, client data, or sensitive media, that is not a theoretical risk. It is a structural problem with cloud-first AI tooling that no privacy policy can fix.
How it solves it#
Three-tier AgentBrain router
Routes every request through Reflex (pattern-matched, under 100ms, no LLM call), Instinct (single-shot, 1 to 3 seconds), or Deliberation (full ReACT loop with 3 to 10 LLM calls). Tier 2 can escalate to Tier 3 mid-response. Simple commands return instantly while complex multi-step tasks get the reasoning depth they need.
Autonomous screen agents on a real desktop
Agents run on a real Ubuntu/XFCE desktop (Xvfb at 1024x1024), controlled by a Gemma 4 vision model with closed-loop servo targeting: ballistic move, single correction with crosshair overlay, then zoom-cropped analysis around the cursor. A live VNC viewer streams the agent's screen into the chat UI in real time.
Multi-agent coding swarms
Up to 20 parallel coding agents run in isolated git worktrees using Claude Code or fully local Cline via Ollama. The swarm planner handles dependency-ordered merging and tracks cost per run. Flight Mode lets the whole swarm operate without internet when running local models.
Local video and media pipeline
Runs Wan 2.2 (5B and 14B MoE), CogVideoX-5B, and LTX-2.3/2.5 for text-to-video and image-to-video locally. Includes an Audio Foundry plugin for ACE-Step 3.5B music generation, neural voice cloning with a consent gate, and a built-in video editor with a 3-lane timeline and ffmpeg text overlays.
Hybrid RAG with Autoresearch
Combines BM25 and vector retrieval over your documents using LlamaIndex, with AST-aware code chunking, entity extraction, and per-project isolation. RAG Autoresearch tunes retrieval parameters overnight and shows chunk scores and sources in the chat UI.
Self-improvement engine with guardian review
A test-fix-verify loop proposes code patches, queues them for guardian review (optionally routed through an Anthropic API check), and broadcasts accepted fixes to connected machines. A codebase lock toggle and pending fixes queue let you stage and approve every change before it applies.
Strengths and trade-offs#
Strengths
- No per-token fees on any workloadEvery inference runs on your GPU using local models via Ollama or Diffusers. Running 20 agents in parallel overnight costs only electricity. LangSmith, Azure AI Studio, and Vertex AI Agent Builder all bill by token or API call; the cost of a long swarm run on cloud platforms scales linearly with work, while Guaardvark's marginal cost is zero.
- Screen agents with a real Ubuntu desktopUnlike sandboxed browser-only agents on commercial platforms, Guaardvark screen agents run on a real Xvfb/XFCE desktop that vision models recognize instantly because it is standard Ubuntu. Agents can open the file manager, interact with the taskbar, and run any application, not just a browser.
- MIT license with no open-core splitThe full platform is MIT licensed, including the agent engine, media pipeline, and GPU orchestrator. You can fork it, modify it, and use it commercially with no restrictions. There is no enterprise tier gating capabilities that are present in the self-hosted build.
- Fully offline Flight Mode, tested end-to-endFlight Mode is a real tested mode, not a marketing label. Swarm tasks, agent runs, and media generation all work without internet when running local models. Automatic network detection triggers local-model fallback without manual intervention.
Trade-offs
- -GPU required for video and high-throughput tasksVideo generation with Wan 2.2 14B MoE and 4K/8K upscaling (Real-ESRGAN, HAT-L) need a capable GPU with sufficient VRAM. Version 2.7.0 introduced 16GB-native video. Running image generation, video, and a coding swarm simultaneously requires the GPU orchestrator to queue tasks; teams with CPU-only hardware will not get the full experience.
- -Small plugin ecosystem at launchGuaardvark launched in March 2026 and ships 10 first-party plugins, compared to the much larger ecosystems of ComfyUI or OpenWebUI. If a specific plugin you need does not exist yet, you will need to build it or wait. The bidirectional MCP interface partially bridges this gap for tools in Claude Desktop and Cursor.
- -Self-hosting is the only pathThere is no managed cloud option, by design. Every user manages their own install, updates, and backups. The one-command curl installer handles initial setup, but ongoing maintenance (GPU driver updates, model management, backup and restore) is on the operator. Teams without Linux server experience should plan for meaningful setup time.
guaardvark vs alternatives#
Guaardvark vs LangSmith
LangSmith is a hosted platform for tracing, evaluating, and debugging LLM runs. Its focus is observability: logging every agent step, replaying runs, and comparing evaluation datasets on a managed dashboard.
| Feature | Guaardvark | LangSmith |
|---|---|---|
| License | MIT | Proprietary |
| Deployment | Your hardware | Cloud-hosted |
| Agent execution | Full (models, tools, screen agents, swarms) | Observability layer only |
| Local models | Yes, via Ollama | No |
| Offline operation | Yes (Flight Mode) | No |
Guaardvark is the better choice when you need to run agents locally with no external API calls, or when you want the full execution stack with built-in models and swarms. LangSmith is still the better fit when you need structured observability across a team (run comparisons, annotation queues, dataset management) without managing your own infrastructure. The two are not direct competitors: LangSmith sits above the execution layer; Guaardvark is the execution layer.
Guaardvark vs Google Cloud Vertex AI Agent Builder
Vertex AI Agent Builder is Google's managed service for deploying AI agents on Google Cloud, with enterprise SLAs, IAM, and audit logging.
| Feature | Guaardvark | Vertex AI Agent Builder |
|---|---|---|
| License | MIT | Proprietary |
| Hosting | Your hardware | Google Cloud |
| Per-token fees | None | Yes |
| Screen agents | Yes (real Ubuntu desktop) | No |
| Local video generation | Yes (Wan 2.2, CogVideoX) | No |
| Offline operation | Yes | No |
Guaardvark wins on cost for high-volume local inference, on data sovereignty for teams that cannot send data to Google, and on capabilities like screen agents and local video that Vertex AI does not offer. Vertex AI Agent Builder is better when you need Google's managed scaling, its compliance certifications, or deep integration with BigQuery and Cloud Storage.
Guaardvark vs Microsoft Azure AI Studio
Azure AI Studio is Microsoft's managed environment for building and deploying AI models and agents on Azure.
| Feature | Guaardvark | Azure AI Studio |
|---|---|---|
| License | MIT | Proprietary |
| Hosting | Your hardware | Azure cloud |
| Per-token fees | None | Yes |
| Custom local models | Any GGUF or LoRA | Azure-hosted models |
| Multi-agent swarms | Up to 20 parallel, local git worktrees | Framework-dependent |
| Offline operation | Yes | No |
Guaardvark is better when you need custom GGUF models, local LoRAs, or data that must stay off Azure. Azure AI Studio is better when your organization runs on Azure, needs enterprise SLAs and Active Directory integration, or when managed deployment justifies the per-token cost.
Install and self-host#
Self-hosting runs via a one-command install script that downloads dependencies, runs migrations, and starts the web UI on port 5173 and the API on port 5000.
```bash
curl -fsSL https://guaardvark.com/install.sh | bash
```What it's built on#
- Languages
- JavaScriptPython
- Frameworks
- FlaskReact
- Databases
- PostgreSQL
- Cache
- Redis
FAQ#
Does Guaardvark require a cloud account or internet connection?
No. Guaardvark runs entirely on your hardware using local models via Ollama. Flight Mode is a tested first-class mode that disables all outbound traffic. The Interconnector for multi-machine clusters and the optional Anthropic API guardian for self-improvement review are the only components that touch the internet, and both are opt-in.
What GPU do I need to run Guaardvark?
The platform runs on consumer GPUs. Version 2.7.0 introduced 16GB-native video generation. For image generation and agent tasks, 8GB VRAM is workable. The GPU Memory Orchestrator schedules tasks to avoid VRAM contention when running multiple models simultaneously, so you can queue video generation while a coding swarm is running.
Is Guaardvark a replacement for LangSmith?
Partially. LangSmith focuses on agent observability: tracing, logging, and evaluation of LLM runs. Guaardvark is a full agent execution environment with built-in models and tooling, not just an observability layer. If you need hosted tracing with team dashboards and no infrastructure to manage, LangSmith is still the better fit. If you want to run agents locally with no external calls, Guaardvark covers the execution side.
How does the self-improvement engine work?
The engine runs a test-fix-verify loop where it proposes code patches based on observed failures or thumbs-up distillation from saved interactions. Each proposed fix goes to a pending queue where you review it before it applies. An optional mode sends the proposed fix to an Anthropic API guardian for a second check. A codebase lock toggle halts the engine entirely if you want full manual control.
Can I use Guaardvark with Claude Desktop or Cursor?
Yes. Guaardvark ships an MCP server that connects to Claude Desktop, Cursor, and other MCP-compatible clients. It also acts as an MCP client, so it can call tools from those environments. The MCP server uses a strong default-deny policy: desktop control, agent execution, and browser automation tools are hidden by default and must be explicitly allowlisted.
Similar open-source tools#
RuFlo
Deploy intelligent AI agents with ease.
thoughtdag
Make LLM context visible, editable, and graph-structured
maka
Local-first AI agent workspace with append-only execution logs
Open Agents
Spawn coding agents that run infinitely in parallel cloud VMs
Claudoscope
Free macOS app for browsing and managing Claude Code sessions
Breadcrumb
Open source LLM tracing and monitoring for AI agents