Heym
A self-hosted visual platform for building, running, and inspecting AI workflows.
- Source repo
- heymrun/heym
- Stars
- ★ 1.3k
- Last updated
- today
- License
- NOASSERTION
- Primary language
- Python
- FA score
- 46/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformCodex · Claude.ai
- You'll need
- Typical use
- An operations team needs to connect webhook input, HTTP requests, Slack notifications, and Redis or RabbitMQ handling in a visual workflow it can host itself.
- Main limitation
- Operation requires Docker, PostgreSQL, Python, Bun, and UV, so it carries more infrastructure responsibility than a hosted automation service.
- Source review
- 46/100 · Major gaps
What does this agent do, and when should you use it?
Heym is a self-hosted AI-native workflow automation platform built with a Vue 3/TypeScript/Vite/Bun frontend and a Python/FastAPI/SQLAlchemy/PostgreSQL backend. Its Vue Flow canvas connects LLM, AI Agent, RAG, HTTP, message-queue, Playwright, and logic nodes, with independent graph branches executing concurrently. A workflow can run from the canvas or be delivered through REST execution endpoints, SSE streaming, an MCP server, or a `/chat/{slug}` Portal interface. Execution history, LLM traces, evals, logs, cost analytics, and OpenTelemetry export support inspection of live and completed runs. Deployment remains on the operator's infrastructure, with skills and Codex workspaces executed through isolated sibling containers as described in the README.
Heym accepts inputs from nodes such as Input (Webhook), Cron, and RabbitMQ Receive, then follows the canvas graph to run LLM, AI Agent, Qdrant RAG, HTTP, Slack, Redis, RabbitMQ, Grist, Drive, Crawler, and Playwright nodes. Condition, Switch, Loop, Merge, and Error Handler nodes determine routing, synchronization, and failure behavior. AI Agent nodes can call tools, canvas-node tools, sandboxed Python tools, MCP connections, and skills packaged as SKILL.md plus optional Python files. Finished workflows can be invoked through /execute or /execute/stream, exposed as MCP tools at /api/mcp/sse, or published as streaming chat applications at /chat/{slug}. The platform records runs, debug logs, request/response payloads, tool calls, token counts, USD costs, and OTLP/HTTP spans for workflows, nodes, and Agent tools.
- An operations team needs to connect webhook input, HTTP requests, Slack notifications, and Redis or RabbitMQ handling in a visual workflow it can host itself.
- An AI product team needs to search managed vector-store content from PDFs, Markdown, CSV, or JSON before passing retrieved context to an LLM or Agent.
- A business process owner needs an Agent to pause for approval, clarification, or additional input before the workflow resumes.
- An internal-tools team wants to publish a workflow as a
/chat/{slug}interface with streaming output, file uploads, and optional authentication. - A team using Claude Desktop, Cursor, or another MCP client wants to expose completed workflows as callable MCP tools.
- A team automating browser interactions needs Playwright steps and runtime selector recovery through Auto Heal.
How do you install or deploy this agent?
Prerequisites listed by the repository are Docker, Python 3.11+, Bun 1.0+, and UV. Clone and start locally:
git clone https://github.com/heymrun/heym.git && cd heym && ./run.shThen open port 4017 in a browser. For environment-based configuration, run cp .env.example .env and then ./run.sh; the script generates SECRET_KEY and ENCRYPTION_KEY. For production deployment, run cp .env.example .env followed by ./deploy.sh.
How do you use this agent?
Open the editor on port 4017, add and connect Input, LLM or AI Agent, and Output nodes on the canvas, then run the workflow from the canvas. External services can call the documented /execute or /execute/stream endpoints. To make a workflow available to MCP clients, expose it through /api/mcp/sse; to provide a chat interface, publish it at /chat/{slug}. Configure an Agent node with MCP connections or upload/create a skill containing SKILL.md when it needs external tools or reusable instructions.
What are this agent's strengths and limitations?
- The same workflow can be delivered through the canvas, REST, SSE, MCP, and a Portal chat UI without rebuilding its logic for each channel.
- Independent DAG branches execute in parallel and can be synchronized explicitly with a Merge node.
- Built-in execution history, LLM traces, evals, token/USD cost tracking, and OpenTelemetry spans provide operational inspection in one product.
- The Agent node combines tool calling, sandboxed Python, MCP, skills, and optional persistent memory in a visual workflow step.
- Operation requires Docker, PostgreSQL, Python, Bun, and UV, so it carries more infrastructure responsibility than a hosted automation service.
- The README warns that mounting the Docker socket grants broad host control; MCP stdio servers require a Docker daemon and fail closed without one.
- Per-run sandbox isolation requires Docker Engine 25.0+, while skills and the Codex node depend on a shared
heym-codex-workspacesvolume. - Licensing needs review before adoption: repository metadata says
NOASSERTION, while the README describes MIT with a Commons Clause that restricts selling or offering the software as a paid service.
How does this agent compare with similar options?
The README compares Heym with n8n, Zapier, and Make.com, highlighting claimed differences including LLM Batch API status branches, persistent Agent memory graphs, Workflow Analyzer, workflow-powered dashboards, an agentic Kanban board, skills, Playwright Auto Heal, native OpenTelemetry export, and per-model USD cost tracking. These are repository claims; teams should validate required integrations and licensing before migrating.
Key facts side by side with the most closely related agents.
| Agent | Source review | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|
| Heym This agent | 46 · Major gaps | ★ 1.3k | today | Python | Codex · Claude.ai |
| Auto Browser | 62 · Some gaps | ★ 793 | 1mo ago | Python | OpenAI API · Claude API |
| Flow-Like | 55 · Major gaps | ★ 955 | 5d ago | Rust | — |
| ByteChef | 53 · Major gaps | ★ 1k | today | Java | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence: README mentions HITL and guardrails but lacks implementation details; SECURITY.md discloses multiple fixed vulnerabilities (e.g., sandbox escape, SSRF, RCE) indicating a security response process, but no specific evidence for least privilege, user confirmation, data flow transparency, sensitive data handling, dependency security, external effects, rollback, or source attribution. Deductions: these aspects are only partially mentioned or lack detailed evidence.
Evidence: Test files (e.g., test_active_execution_registry_resilience.py) demonstrate resilience to database failures and cover failure messages and degradation paths; but no evidence for dependency availability (e.g., lock files, images). Deductions: insufficient evidence for dependency availability.
Evidence: README describes multiple use cases (e.g., multi-agent orchestration, RAG, MCP) and provides quick start and deployment instructions; but capability boundaries (e.g., limitations, unsupported features) and trigger precision (e.g., event trigger conditions) are not clearly defined. Deductions: insufficient evidence for capability boundaries and trigger precision.
Evidence: README provides installation instructions, examples, and feature lists, but lacks explicit version changelog, known limitations, and naming stability; license is MIT + Commons Clause but metadata is NOASSERTION. Deductions: incomplete version changelog, known limitations, naming stability, and license metadata.
Evidence: README showcases rich features (e.g., visual editor, multi-agent orchestration, observability), but no cost-benefit analysis or quantitative comparison with other products. Deductions: insufficient evidence for cost-benefit.
Evidence: Feature claims in README lack traceable test or documentation links; test files exist but do not cover all claims; facts and inferences are not clearly separated. Deductions: insufficient claim traceability, cross-source corroboration, and fact-inference separation.
- License metadata is NOASSERTION, actual is MIT + Commons Clause, need to confirm compliance.
- SECURITY.md discloses multiple fixed vulnerabilities, but no current security audit report is provided.
- Feature claims in README lack verifiable test or documentation support.
FAQ
Can I run Heym on my own infrastructure?
./run.sh, ./deploy.sh, and Docker-based deployment paths and positions the product as self-hosted.How can Claude or another AI client call a workflow?
/api/mcp/sse; the README explicitly names Claude Desktop, Cursor, and other MCP clients.