boommanpro/gaia-workflow-engine
boommanpro/gaia-workflow-engine has 223 GitHub stars and 21 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-15 23:57 UTC. Last push 2026-11-11.
Measured star growth
- in 24h
- 0
- in 7 days
- Collecting history
- in 15 days
- Collecting history
- in 1 month
- Collecting history
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
一个现代化的可视化规则引擎平台,用于编排复杂的 AI 工作流。在无限画布上设计、测试和部署 AI 流程——无需编写代码。结合 flowgram.ai 的能力与 Java 服务端,提供生产级工作流管理。在此之上提供AI Agent / Copilot的助手编排能力。
AIREADME.en.md
OriginalGaia
Visual AI Workflow Editor
Built on flowgram.ai · React + Spring Boot + SQLite
简体中文 · Release Log · Issues · Discussions
Arrange AI workflows on an infinite canvas with drag and drop: LLM calls, conditional branches, loops, HTTP requests, code execution. Or describe what you need in the sidebar chat and let the AI plan the steps, create the nodes and edges, and run the workflow on the canvas right away.
The backend is a Spring Boot monolith with built-in SQLite storage and the AgentScope execution engine, packaged as a single executable jar.
Demo
A 76-second end-to-end run: describe the task in chat, let the AI generate the workflow, unit-test nodes, then run the whole flow. Everything is recorded from real API calls.
https://github.com/user-attachments/assets/2530ba1e-3ac2-49c3-835f-87ed9f77b44c
Timeline
| Time | What happens |
|---|---|
| 0:00 | Canvas overview: Start → Condition → Copywriting → End |
| 0:18 | Sidebar chat: describe the task, the AI plans five steps |
| 0:46 | Auto orchestration: nodes created and wired, three unit tests pass (sped-up replay) |
| 0:55 | Trajectory: every tool call and execution step is traceable |
| 1:01 | Test run: input 85, the "pass" branch fires and returns the verdict |
Try it live: Home · Admin Console · Release Log
Features
Canvas & Nodes
- Infinite zoomable canvas with drag-and-drop editing; loops, branches, and conditional logic
- Multi-branch node with and/or condition groups and an else fallback
- Built-in nodes: LLM call, code execution, HTTP request, string formatting, variable management
- Per-node unit tests and full-workflow runs, with live input/output and execution history
- Canvas plugins: minimap, auto-layout, snap lines
Chat Generation
- Generate workflows from the sidebar chat: CreatePlan step-by-step planning, Canvas tool creates nodes and edges
- HTTP / Query nodes can run lookups during the conversation
- Live preview of token usage and plan progress
- Sessions are renameable; every Trajectory is fully traceable
Configuration & Operations
- Agent config center: model parameters, system prompt, knowledge base, tool definitions
- Switchable execution engine: built-in AgentScope, or Volcengine Ark Managed Agents
- Every save creates a version, with diff between versions and one-click switch
- Workflow templates: spin up new workflows from a template
- Observability API (
/api/agent/metrics/*): tool calls, session runs, LLM call ledger, context usage stats
Quick Start
Prerequisites
| Dependency | Version |
|---|---|
| Node.js | >= 18 |
| JDK | >= 21 (code nodes are compiled at runtime with -source 21; run the service on JDK 21) |
| Maven | 3.x |
The project itself builds for Java 17 (
maven.compiler.release=17), but code nodes are compiled on the fly by the running service through the system JDK, so the runtime requires JDK 21 or above.
Frontend
git clone https://github.com/boommanpro/gaia-workflow-engine.git
cd gaia-workflow-engine/apps/console
npm install
npm run dev
Then visit http://localhost:3000.
Backend
cd apps/api/gaia-workflow
mvn install -DskipTests # required on fresh clone: installs base/infra artifacts locally
cd gaia-workflow-app
mvn spring-boot:run # dev mode, port 48080
Multi-module note: running
gaia-workflow-appalone resolves base/infraSNAPSHOTartifacts from the local Maven repository. After cloning (or whenever base/infra change), runmvn install -DskipTestsfrom the aggregator directoryapps/api/gaia-workflowfirst.
For production, package an executable jar:
cd apps/api/gaia-workflow
mvn package -DskipTests
java -jar gaia-workflow-app/target/gaia-workflow-app-0.0.1-SNAPSHOT.jar
The API server runs on port 48080. The SQLite database file gaia_workflow.db is created in the process working directory; first startup initializes the schema and seeds default configuration automatically.
LLM Configuration (required)
Agent chat requires an OpenAI-compatible LLM endpoint (vLLM / Ollama / LM Studio / any gateway). After startup, open Admin → Config → Model Settings in the browser, fill in API Host / API Key / model name, and save. The config is stored in the agent_config table and takes effect immediately without restart.
- The default engine is the built-in AgentScope (Harness ReAct loop; depends on io.agentscope:agentscope-harness, publicly available on Maven Central).
- Alternatively switch to the Volcengine Ark Managed Agents engine: fill in apiKey / environmentId / model ID in the "Ark Managed" section of the Model Settings page, then flip the default engine to Ark.
Docker Compose deployment
Prerequisite: Docker Engine with Docker Compose v2. Copy the non-secret example configuration, then build and start the stack:
cp .env.example .env
docker compose up --build -d
On Windows PowerShell, use Copy-Item .env.example .env. Open http://localhost:8080; the health endpoint is http://localhost:8080/api/health. Only the Nginx web service publishes a host port. api:48080 is reachable only on the Compose network. After the first start, configure an OpenAI-compatible model endpoint, model name, and API key in Admin → Config → Model Settings. Do not place those credentials in .env.example or commit them.
Useful operations:
# Status and logs
docker compose ps
docker compose logs -f web api
# Stop while retaining the named SQLite volume
docker compose down
# Rebuild after a source or image update
docker compose up --build -d
# Delete all local SQLite data only when that is intentional
docker compose down -v
The SQLite file is /data/gaia_workflow.db in the api container and is persisted by the named gaia_data volume. A normal docker compose down does not remove it; after recreating containers, verify it with docker compose exec api ls -l /data/gaia_workflow.db. SQLite supports one API writer only: do not horizontally scale this Compose deployment to multiple API instances that write to the same database.
The Docker frontend image injects same-origin /api at build time without changing the existing backend-address default for standalone or desktop deployments. Nginx preserves the /api/ prefix and proxies API/SSE traffic with HTTP/1.1, buffering and caching disabled, so no permissive CORS policy is required. It also forwards /mcp/ for managed-Agent callbacks and falls back to index.html for SPA routes.
Built-in vs. managed Agents
Local Compose supports the built-in Agent path by default; it does not make a managed Agent work out of the box. host.docker.internal:48080 is not a production callback address and is not the default. If an externally hosted execution engine (such as AgentScope Service) must call Gaia's MCP endpoint, the deployer must supply a publicly reachable HTTPS base URL:
GAIA_MCP_PUBLIC_URL=https://gaia.example.com
The public reverse proxy or ingress must route /mcp/ (and /api/ for the UI) at that address to this stack's Nginx web service. Leave the value empty when Compose is local and only the built-in Agent is used. If a managed-Agent provider configuration has already been saved in the admin UI, its mcpPublicUrl overrides the application setting and must be updated to the same HTTPS address.
Troubleshooting:
- Frontend 502: run
docker compose ps, wait forapito becomehealthy, then inspectdocker compose logs apifor startup or migration errors. - SSE disconnects: use the Nginx
webentrypoint and ensure any upstream proxy does not buffer responses or impose a short read timeout. The supplied/api/proxy disables buffering and uses a one-hour timeout. - SQLite permission errors: do not replace the ownership of mounted
/data; inspect it withdocker compose exec api ls -ld /data. The image runs as the non-rootgaiauser. - Missing model configuration: save the OpenAI-compatible endpoint, model name, and API key in Admin model settings. They are persisted in SQLite and must not be added to the image or example environment file.
Tech Stack
| Layer | Technology |
|---|---|
| Frontend | React 18 · TypeScript · flowgram.ai 1.0.12 |
| Backend | Spring Boot · MyBatis Plus · Java 17 (build) / 21 (runtime) |
| Database | SQLite |
| Build | rsbuild (frontend) · Maven (backend) |
Deployment Options
- Separate frontend/backend deployment, for environments with an existing gateway and service mesh
- Single-jar all-in-one deployment, for small to medium deployments
- Frontend on GitHub Pages with a configurable backend address
- Containers: official Docker Compose deployment (Nginx frontend entrypoint + Spring Boot API)
Documentation
- Project docs: https://boommanpro.github.io/gaia-workflow-engine/#docs
- flowgram.ai framework docs: https://flowgram.ai/guide/
Contributing
Issues and PRs are welcome. Commit messages follow Conventional Commits:
- Fork the repo and create a feature branch
- Commit your changes and open a Pull Request