Open WebUI Tools Collection
A modular suite of Open WebUI tools, function pipes, and filters that turns your instance into an AI workstation spanning academic research, image/music/video generation, and autonomous agents.
Evidence shows tools configure API keys via Valves and call external services; the README does not document least-privilege network scope, user confirmation flows, data-flow mapping, or rollback. The actual tool source is absent from the evidence set, so sensitive-data handling and dependency pinning cannot be verified — all scored low. SECURITY.md provides a private reporting channel, but this does not substantiate implementation details. Deductions: no rollback, no default permission statement, no confirmation flow.
Unit tests for two of 30+ components (Atlas Cloud, MiniMax) cover defaults, clamping, and explicit failure messages with error-status emission; failure handling in those files is genuinely careful. However, only 2 components carry test evidence. Deduction: the vast majority of tools have no corroborating tests.
The README organizes tools by scenario (academic, creative, agentic) with prerequisites (ComfyUI, Perplexica, Mopidy, API keys), giving decent audience and environment fit; but trigger precision and capability boundaries for individual tools are one-liners, and boundaries of the Planner Agent v3 are undocumented. Deduction: trigger and boundary documentation is thin.
Good information architecture: TOC, per-tool configuration docs, dual install paths (Hub and manual); MIT license file is complete (copyright Juan Jose Soliz Priore — relation to 'Haervwe' unstated); SECURITY.md honestly states solo-maintainer status with a 48h acknowledgment commitment. Deductions: no CHANGELOG/versioning, no known-limitations section, and several TOC sections (Cloudflare, SearxNG) are truncated in the provided files.
Outputs are designed as markdown image/video links, HTML embeds, and citation metadata, and tests verify output formatting; 20+ tools fill real gaps in search, media generation, and routing for Open WebUI; many tools need no key or only free keys, giving reasonable cost/benefit. Deduction: actual output quality depends on tool source not provided in this static review.
Most README capability claims (real-time status, progress updates) cannot be traced in the provided evidence; only two tools' tests corroborate behavior; documentation does not separate fact from inference, and TOC/body inconsistencies exist. Deductions: weak claim traceability and cross-source corroboration.
- This review is based only on README, LICENSE, SECURITY.md, and two test files; most tool source is absent from the evidence, so all security conclusions are low-confidence static inferences.
- For key-handling tools (YouTube, Pexels, Cloudflare, HF, etc.), manually inspect the corresponding .py source for key storage and transmission before deployment.
- Capability boundaries, failure behavior, and external side effects of autonomous components like Planner Agent v3 are undocumented; validate in an isolated environment first.
- Solo hobbyist maintenance with no versioned releases or changelog; verify commit diffs yourself before updating.
What does this agent do, and when should you use it?
Haervwe/open-webui-tools is an MIT-licensed modular extension collection for Open WebUI (0.6.0+) comprising over 20 specialized Tools, Function Pipes, and Filters. The tool layer covers arXiv paper discovery, Perplexica web search with citations, Pexels/YouTube/SearxNG media search, X (Twitter) data lookup, provider status checks, and image, music (ACE Step 1.5), and video generation (WAN 2.2, Google Veo) via ComfyUI, Hugging Face, Cloudflare Workers AI, and Atlas Cloud backends. The pipe layer includes Planner Agent v3 — an autonomous agent with dependency-aware task decomposition, subagent delegation, asyncio-based parallel execution, MCP server support, and a visual execution tracker — plus arXiv Research MCTS, multi-model conversations, and a Letta Agent integration. The filter layer adds Semantic Router model selection, Prompt Enhancer, and a Doodle Paint canvas. The deployment boundary is explicit: every component installs as a Python file into Open WebUI's Workspace and depends on the Open WebUI runtime plus each tool's declared third-party services and API keys; nothing runs standalone outside Open WebUI.
Components install as .py files into Open WebUI's Workspace > Tools / Functions / Filters. Tool components read user requests through Open WebUI's tool-calling mechanism and hit external APIs: arXiv Search queries arXiv.org directly (no key needed), Perplexica Search calls a self-hosted Perplexica API and emits citations, and the Pexels/YouTube/Xquik/OutageDeck/OpenWeatherMap tools each call their respective APIs and render embedded content in chat. Generation tools submit prompts to ComfyUI's HTTP + WebSocket API (default WAN 2.2 text-to-video workflow, Qwen Edit 2509 image editing, ACE Step 1.5 music), wait for completion, and upload outputs to Open WebUI storage for display via embedded HTML players. The Planner Agent v3 pipe decomposes complex requests into a dependency-aware task tree, concurrently executes tool and subagent calls (web search, image generation, knowledge-base RAG, code interpreter, terminal agent) via asyncio.gather, supports MCP servers with connection deduplication, persists state across turns via attached JSON files, exposes interactive UI modals (plan_approval, ask_user, give_options), and renders a real-time HTML execution tracker. The Semantic Router filter intelligently routes requests to models, while other filters handle document processing, thinking-tag cleanup, and citation handling.
- A researcher who wants to discover arXiv papers and run MCTS-driven deep literature research without leaving an Open WebUI chat
- A content creator who wants to generate and edit images in conversation (Qwen Edit 2509 multi-image style transfer) with inline before/after embeds
- A musician-style user who wants to produce tracks from tags and lyrics with ACE Step 1.5, controlling key, BPM, and language
- A power user automating workflows who wants Planner Agent v3 to autonomously decompose tasks, run search/image/terminal subagents in parallel, and track progress visually
- A self-hoster of Perplexica who wants chat models to answer with real-time, citation-backed web search
- A multi-model user who wants the Semantic Router filter to automatically pick the right model per request
What are this agent's strengths and limitations?
- Unusual breadth in one repository: academic search, web search, image/music/video generation, an autonomous agent, and model routing coexist, avoiding stitching together multiple projects
- Planner Agent v3 offers dependency-aware task trees, asyncio parallel execution, MCP support, and a real-time HTML execution tracker — a notably advanced agent implementation within the Open WebUI ecosystem
- Several components work with zero configuration: arXiv Search and OutageDeck need no API keys, and ComfyUI workflow JSONs ship with the repo, lowering the entry barrier
- Fine-grained VRAM management: most generation tools support optional Ollama model unloading, valuable for GPU-constrained local deployments
- Deeply bound to the Open WebUI runtime (installed as Workspace tools/pipes/filters); components cannot be directly reused on ChatGPT, Claude, or other platforms
- Many core features depend on external services and keys: Perplexica, ComfyUI, and Ollama must be self-hosted, while Pexels/YouTube/Hugging Face/Atlas Cloud/MiniMax/Google each require separate API keys, spreading maintenance overhead
- Parallel execution can cause external race conditions when tools have stateful dependencies within the same turn; the README explicitly warns that complex interdependent workflows need caution and possibly an async database
- Documented limitations exist: Google Veo image-to-video supports only a single input image, and Flux Kontext multi-image support is not yet implemented
How do you install or deploy this agent?
Option 1 (recommended): visit https://openwebui.com/u/Haervwe, browse the collection, and click "Get" on desired tools, following the in-app installation prompts. Option 2 (manual): copy the .py files from the repository's tools/, functions/, or filters/ directories, go to Open WebUI Workspace > Tools / Functions / Filters, paste the code, provide a name and description, and save. Prerequisites: Open WebUI 0.6.0+, Python 3.8+; generation tools require a local ComfyUI instance with the bundled workflow JSONs from the Extras/ folder (e.g., ace_step_api., video_wan2_2_14B_t2v., image_qwen_image_edit_2509_api_owui.); Perplexica components require a self-hosted Perplexica instance.
How do you use this agent?
After installing, enable each tool/pipe/filter in Open WebUI and configure its Valves: Pexels needs PEXELS_API_KEY, YouTube needs YOUTUBE_API_KEY, the Hugging Face generator needs an HF API key, Atlas Cloud needs ATLASCLOUD_API_KEY, the MiniMax pipe needs MINIMAX_API_KEY, and the Veo pipe needs GOOGLE_API_KEY; Perplexica components need perplexica_api_url (default http://localhost:3001) plus chat/embedding models. ComfyUI tools require comfyui_api_url pointing at your ComfyUI server (default http://localhost:8188). Then invoke in natural language, e.g., "Search for recent papers about tree of thought", "Generate an image of a serene mountain landscape at sunset", or "Generate a cyberpunk, darkwave song in E minor, 140 BPM". Planner Agent v3 requires PLANNER_MODEL (a base model ID from Admin Panel > Settings > Models) in its Valves, with optional SUBAGENT_MODELS, ENABLE_PLAN_APPROVAL, and YOLO_MODE.
How does this agent compare with similar options?
The README positions Perplexica as a self-hosted, open-source AI search engine and alternative to Perplexity AI; this repo's Perplexica Search/Pipe builds on it. The generation tools contrast with using AUTOMATIC1111, ComfyUI, or OpenAI DALL·E backends directly — this collection wraps those backends through Open WebUI's native image middleware.