heym

Heym is a self-hosted runtime and platform for building, orchestrating, running, observing, evaluating, and controlling agent systems on users' own infrastructure. According to the official description, it covers multiple stages of the agent lifecycle: building, orchestrating, running, observing, evaluating, controlling, and exposing externally. These stages are interconnected, and observation and evaluation results can be used in turn to optimize prompts, tools, and workflow logic.

Main capabilities

• Building: Supports visual and AI-assisted creation. Workflow graphs can be built and inspected on a canvas, generated or modified by an AI assistant using text or voice, and templates can be reused. Combine LLM and Agent nodes with deterministic logic such as conditions, branches, loops, merges, data transformations, and sub-workflow calls, with support for expressions in supported fields.
• Multi-agent orchestration: Specify named sub-agents and sub-workflows for the orchestrator to call. Independent sub-agent calls can execute concurrently, and reasoning intensity and temperature can be configured per agent.
• Knowledge and memory: Retrieve documents through RAG using Qdrant or PostgreSQL/pgvector; enable a knowledge graph for agents to retain facts across runs, configure sharing permissions between agents, and inspect and edit memory with a graph editor.
• Tools and integrations: Connect communication services, developer tools, productivity apps, databases, queues, and storage through built-in nodes, the HTTP API, and MCP; use the MCP Call node for deterministic tool calls.
• Skills and custom behavior: Attach reusable SKILL.md instructions and optional Python tools to agents, create or modify skills with AI, install plugins, or extend with custom nodes that have typed configuration and execution handlers.
• Running: Supports parallel execution and recovery, with independent nodes running concurrently, using Merge to synchronize branches, and configurable retries, backoff, error paths, and model fallback; can be triggered via cron, webhook, messaging integrations, inbox events, WebSocket messages, and platform events. Supports browser automation with Playwright and selector recovery, as well as running Codex or OpenCode Go as a workflow step against a codebase and producing diffs and pull requests.
• Agent board: Cards are persistent agent tasks. Moving a card into a column runs it according to its ordered workflow chain, combining card content, comments, history, and previous outputs, and writing the results back to the card.
• Load distribution: Distribute eligible back[end tasks] among multiple Heym instances through PostgreSQL.runs on the platform, and execution weight can be configured on the instances page in settings.
• Observability: Open production runs from history or dashboard cards, inspect saved state, and track real-time node progress, debug logs, and final output; view model requests, responses, tool calls, input/output tokens, latency, and errors, calculate USD cost using maintained model prices and custom overrides, and view historical spend by model and time range; track execution volume, success rate, and duration, view logs, set alerts for errors, duration, tokens or USD spend, and run counts, and export workflow, node, and agent tool spans via OpenTelemetry.
• Evaluation and control: Run repeatable prompt evaluations to compare expected and actual output across models and view historical results; the workflow analyzer uses execution context to generate an editable Markdown report explaining purpose, step behavior, and improvement points. Agent checkpoints can pause execution for acceptance, editing, or rejection, then continue from the saved execution snapshot, and review branches can notify people through existing integrations. For security and access, input guardrails, encrypted credentials, team sharing, execution tokens, and workflow authentication can be configured, with support for JWT or OIDC SSO login through providers such as Keycloak, Okta, Entra ID, Auth0, and Google.
• External exposure: Call workflows from applications via REST, SSE, and MCP, stream execution events, or publish workflows as tools available to MCP clients (at /api/mcp/sse). Portals and chat can publish workflow-driven chat interfaces with streaming, uploads, and conversation history; you can also build custom dashboard components backed by workflows, manage structured data tables, and save generated files in Heym Drive with team sharing and share links.

Usage and deployment Try locally: git clone https://github.com/heymrun/heym.git && cd heym && ./run.sh, then open http://localhost:4017 to create an account and start building. run.sh starts PostgreSQL, the backend, and the frontend, and generates SECRET_KEY and ENCRYPTION_KEY when needed. Prerequisites are Bun, Python 3.11+, UV, and Docker. Heym uses PostgreSQL as its core database. The standard Compose setup includes PostgreSQL, the backend, and the frontend, and uses local volumes to store persistent files. RabbitMQ and Redis are integrations connected on demand; RAG can use the built-in PostgreSQL/pgvector backend or an external Qdrant server; S3 is used through its integration node. Production deployment uses Docker Compose and Python 3. The official recommendation is to first register an administrator account, then set ALLOW_REGISTER=false in .env and restart to disable registration. There are also prebuilt Docker images containing the frontend and backend. The example connects it to an existing PostgreSQL database on the host, and DATABASE_URL must be set; when using the host.docker.internal hostname on Linux, add --add-host=host.docker.internal:host-gateway. Mount and runtime settings support Playwright (--shm-size 2g), Heym Drive (FILE_STORAGE_DIR=/app/data/files), plugins (HEYM_PLUGINS_DIR=/app/data/plugins), Codex workspace, OpenCode Go workspace, and Docker access (for sandbox containers and MCP stdio execution). For multi-instance execution, connect more Heym instances to the same PostgreSQL database to distribute eligible background workflow runs. The main instance provides the UI, API, MCP, and portal and owns local files and installed plugins, while worker instances execute eligible queued runs.

Examples and use cases The official demo shows an end-to-end e-commerce sales campaign workflow, including generating a workflow with an AI assistant, human review, dashboard execution, structured data, dashboards, tracing, RAG, MCP, analytics, and team collaboration. There is also a demo of an orchestrator and sub-agent system: for "For requests such as "How to get from Berlin to Frankfurt" and "What should I eat when I get there," the orchestrator can, within the same turn, invoke mutually independent sub-agents, run them concurrently, and merge the results. Workflows can also be generated through natural language, invoked from chat, and skills attached to agents can be created.

Limitations and Status Heym is in a pre-1.0 release series, and the official recommendation is to follow the version badge and release notes for updates; the deployment guide and security policy provide references for operational setup. The above features and components are subject to the text in the official repository. Prices, platform support scope, or effects not mentioned are not in the materials and are not currently provided.