VectorVein
No-code AI workflow software: drag and drop nodes to build automated workflows powered by LLMs and your personal knowledge base.
The project exposes workflow execution via a local FastAPI service (default localhost:8787) and sends user content to configurable LLM endpoints, including an option to upload screen screenshots with voice conversations; no least-privilege design, sensitive-data encryption notes, or pre-execution confirmation mechanisms are visible. No malware or covert exfiltration was found, so not zero. Dependency security is handled reasonably (Dependabot auto-merge that skips major updates and gates on frontend checks), but auto-merge itself carries risk, so 2. Rollback is limited to a database-backup warning with no overall recovery plan.
README and code layout (backend/frontend, pdm, migrations) are self-consistent; version/tag fail-fast validation and provided test/lint/typecheck commands plus test files (conditional branching, embedding settings migration) support basic consistency. Error and failure messaging is barely evidenced beyond the webview2 troubleshooting note, so failure_messages is 1. Dependencies are pinned via PDM/pnpm lockfiles, giving 2.
The README clearly targets non-programmers with concrete everyday scenarios (translation, mind maps, complaint classification), so audience/scenarios is 2; environment requirements (Python 3.8–3.11, webview2, cross-platform packaging) are explicit. Capability boundaries (unsuitable tasks, unreliable LLM output handling) are not documented, hence the deduction.
Information architecture is good (installation, configuration, usage, development sections, multilingual README, tutorials); install notes are concrete; examples/FAQ covered by templates and docs; versioning has tag validation and automated releases. However, the LICENSE is a custom non-standard license (non-commercial restriction, share-alike, PRC governing law) with GitHub metadata NOASSERTION, creating legal uncertainty for enterprise adoption, so license is 1; there is no known-limitations section; maintenance responsibility rests on one unverified publisher with only a contact email and no governance statement.
Output usability is supported by run records, dual user/editor interfaces, and concrete outputs (Word/Excel/mind maps). Marginal value comes from the no-code + local knowledge base + local LLM (Ollama/LM-Studio) combination with its own positioning versus langflow. Cost-benefit is not discussed (API costs, token usage, local resource needs), so 1.
README claims (API endpoints, embedding backend migration) are partially corroborated by tests (test_embedding_settings.py, test_conditional_branching.py), but coverage is narrow and most claims cannot be traced; this is a static review with no execution, and cross-source corroboration is only partial; facts and inference are mostly separable, but several README assertions (quality checks, automated builds) cannot be verified from the given files.
- The license is a custom non-standard license (non-commercial restriction, share-alike for derivatives, PRC governing law) with GitHub metadata NOASSERTION; obtain legal review before commercial use.
- The local API service starts automatically and can execute workflows with no mention of authentication; other processes on the same machine may invoke it — verify port exposure before deployment.
- Voice conversations can automatically capture and upload screen screenshots, a sensitive-data egress path; confirm where the content goes before enabling it.
- The dependency auto-merge workflow carries supply-chain risk; consider manual review even for minor updates.
- Back up the database before schema changes as the README warns, or data may be lost.
- The publisher is an unverified individual/team with no evidence of governance or long-term maintenance commitment.
What does this agent do, and when should you use it?
VectorVein (GitHub: AndersonBY/vector-vein) is a no-code AI workflow software inspired by LangChain and langflow. Built on a Python backend (pywebview) with a Vue3 + Vite frontend, it ships as a desktop application where users design workflows by dragging nodes onto a canvas and wiring them in the editor, then run them from the user interface. It supports OpenAI-compatible API endpoints, connecting to local services like LM-Studio, Ollama, and vLLM, plus remote and custom LLM configurations. Since v0.4.0 it embeds a local FastAPI service (default port 8787) for programmatic workflow listing and execution. Workflows can consume Word documents, web URLs, and tables, and produce translated documents, mind maps, and classification spreadsheets, with vector embedding search and a speech-driven Agent on top.
VectorVein lets users compose workflows by connecting input, processing, and output nodes in a visual editor. At run time it invokes the configured large language models (OpenAI-compatible endpoints, remote LLMs, or custom LLMs) to perform tasks such as translating Word documents, summarizing web articles into Chinese mind maps, and auto-classifying customer complaints into an Excel output. It configures embedding models via the vv-llm embedding_backends scheme for vector search, supports OpenAI-compatible speech recognition (e.g., Groq), and offers shortcuts that launch voice conversations with the Agent (optionally attaching a screen screenshot). Since v0.4.0, a local FastAPI server starts automatically with the app, exposing GET /api/workflow/list, POST /api/workflow/run, POST /api/workflow/check-status and more, accepting a workflow id and input_fields for parameterized execution; interactive docs live at http://localhost:8787/docs. Official templates and an online tutorial help users get started.
- Office workers who regularly translate English Word documents: build a translation workflow that takes an English document as input and outputs the translated document.
- Researchers digesting English material: use the web article summary workflow to turn an article URL into a Chinese mind map.
- Customer service managers: upload a table of complaint content with custom keywords and get an auto-generated Excel with classification results.
- Users who want LLMs in a private local environment: connect via OpenAI-compatible endpoints to LM-Studio, Ollama, or vLLM so inference stays on the machine.
- Developers integrating AI flows into other systems: call the local REST API on port 8787 to run workflows programmatically and poll their status.
- Voice-first users: configure a speech recognition service and shortcuts to start talking with the Agent in one keystroke.
What are this agent's strengths and limitations?
- Genuinely no-code: drag-and-drop node wiring builds workflows, with official templates and an online tutorial lowering the learning curve.
- Flexible model access: since v0.2.10 API endpoints are separated from LLM configs, allowing multiple endpoints per model and OpenAI-compatible connections to LM-Studio, Ollama, and vLLM for fully local operation.
- A local REST API (default port 8787, with OpenAPI docs) makes it easy to integrate workflows into other apps and automation tools.
- Covers multimodal scenarios: beyond text workflows, it supports vector embedding search, speech recognition with voice-shortcut conversations, and local Stable Diffusion integration.
- License is NOASSERTION and the README states personal non-commercial use only — commercial adoption requires checking LICENSE.md yourself.
- Depends on pywebview/webview2: Windows users must install the webview2 runtime, and a locked zip archive will prevent the app from opening.
- Desktop-application shape aimed at personal local use; the repository documents no cloud deployment or multi-user collaboration story.
- Speech recognition and embedding features rely on external OpenAI-compatible services, which require extra setup when self-hosted.
How do you install or deploy this agent?
1) Download the package from GitHub Releases (https://github.com/AndersonBY/vector-vein/releases/) and unzip it; if it won't open, right-click the zip and choose 'Unblock'. 2) The app is built on pywebview with the webview2 kernel; on Windows install the webview2 runtime (https://developer.microsoft.com/en-us/microsoft-edge/webview2/). 3) On first run, a 'data' folder is created in the installation directory for the database and static resources. 4) Open the app and configure at least one usable LLM API endpoint in settings (OpenAI-compatible interfaces, e.g., the OpenAI API or local LM-Studio at http://localhost:1234/v1/ and Ollama at http://localhost:11434/v1/). For source development: the backend needs Python 3.8–3.11 and PDM — run pdm install in the backend directory (pdm install -G mac on Mac), then pdm run dev; the frontend needs pnpm install, and backend resources are packed with pdm run build-front; package an executable with pdm run build. No programming is required to use the app.
How do you use this agent?
After launching, add official templates to learn the workflow system or follow the online interactive tutorial (https://vectorvein.ai/workspace/workflow/editor/tutorial). In the editor, name the workflow, add tags, drag nodes from the left node list onto the canvas, and wire them into a flow; for daily use, switch to the user interface, enter input, click the run button, and review history under 'Workflow Run Records'. To invoke programmatically: requests.post('http://localhost:8787/api/workflow/run', ={'wid': 'your-workflow-id', 'input_fields': [{'node_id': 'node1', 'field_name': 'input', 'value': 'Hello World'}], 'wait_for_completion': True}), with POST /api/workflow/check-status for status and full docs at http://localhost:8787/docs. Configure embedding backends in the Embedding models settings for vector search; to use a local Stable Diffusion API, set set COMMANDLINE_ARGS=--api in webui-user.bat.