The team behind Heym
Two engineers building the AI workflow platform they always needed — source-available, self-hosted, and designed for production from day one.
Why we built Heym
We built Heym because the tools we reached for kept getting in the way. We wanted to build AI workflows — multi-step pipelines where LLMs reason, call tools, and adapt their behavior based on context. But existing automation platforms were designed for deterministic, rule-based logic. Wiring an LLM into n8n or Zapier meant fighting the platform on every node.
We decided to build what we actually wanted to use: a workflow platform designed from the ground up for AI-native workloads. Visual canvas because that's how we think about agent architecture. MCP support because tool calling is the primitive that matters. Self-hosted because the workflows we needed to build involved data that couldn't leave our infrastructure. Source-available because the ecosystem moves faster when everyone can inspect, extend, and contribute.
Every decision in Heym reflects a constraint we hit in practice. The Agent node's automatic context compression exists because we ran agents on long documents and watched them silently fail as they hit the model window. Human-in-the-loop review exists because some decisions — a drafted email, a generated report, a data transformation — need a human sign-off before they proceed. The Evals tab exists because we got tired of manually checking whether a prompt change broke something downstream.
Heym is the platform we needed and couldn't find. We're two engineers building it in public — shipping features, listening to users, and treating the gap between AI research and production deployment as the real problem worth solving. If you're building AI-powered systems and want a tool that gets out of your way, Heym is for you.
What we've built
Heym ships as a complete platform for building, running, and evaluating AI workflows. Below is a summary of the capabilities we use ourselves and ship to production every week.
Visual workflow canvas
Build AI workflows by connecting nodes on a drag-and-drop canvas — no code required for most use cases. The editor supports real-time parallel execution, a debug panel with live node outputs, data pinning for testing downstream nodes without re-running upstream ones, and an AI Assistant that generates or modifies workflows from a natural-language description.
Agent node with tool calling
The Agent node is an LLM node with a full tool-calling loop. Agents can run Python tools, connect to external MCP servers, use skill files (instruction content and Python scripts), delegate tasks to sub-agents, and call other workflows as tools. The orchestrator automatically compresses message history when context usage approaches 80% of the model window — keeping the system prompt, first user message, and most recent user message intact — so long-running agents stay reliable without manual intervention.
Human-in-the-loop review
Agent nodes can pause at any checkpoint, generate a public one-time review URL, and wait for a non-logged-in reviewer to accept, edit, or refuse the output before the run continues. A dedicated review output handle lets you send Slack messages, emails, or trigger other notification flows while the run is pending. The same agent run can pause more than once; each decision resumes execution from the exact stored snapshot.
RAG and vector search
The Qdrant RAG node inserts documents into or searches a vector store for retrieval-augmented generation. Manage vector stores from the Vectorstores tab — upload PDFs and text files, then search with metadata filters and optional Cohere reranking. Results flow directly into LLM or Agent nodes with score, text, and metadata attached.
Evals and LLM traces
The Evals tab lets you create test suites, generate test cases, and run evaluations against Agent workflows across multiple models simultaneously. Choose scoring methods — Exact Match, Contains, or LLM-as-Judge with a configurable judge model — and compare pass rates and outputs side by side. The Traces tab shows full LLM execution traces including request/response payloads, timing breakdowns, and every tool call with its arguments and result, making it straightforward to diagnose unexpected agent behavior in production.
Portal and MCP integration
The Portal exposes any workflow as a public chat UI at a configurable slug — no Heym login required for end users. Use it for internal tools, customer-facing chatbots, or AI-powered forms, with optional per-portal auth and file upload support. The MCP tab turns Agent workflow tools into an MCP-compatible endpoint that connects directly to Claude Desktop, Cursor, or any other MCP client with a single JSON config.
Built for production
Heym runs on your own infrastructure with a single docker-compose up. The stack is three containers — PostgreSQL, a FastAPI backend, and a Vite-built frontend — with all credentials encrypted at rest using AES-256. Workflow webhooks support JWT and custom header auth; rate limiting applies to login and portal endpoints out of the box.
Teams and credential sharing let you work collaboratively without sharing raw API keys. Share a workflow, credential, vector store, or global variable with a team and every member gains access instantly. The Global Variable Store persists key-value state across executions — useful for counters, accumulated results, and auth state that agents need across runs.
The execution engine runs nodes in parallel when they have no dependencies on each other. It builds a DAG from the workflow graph, schedules nodes concurrently with a thread pool, and streams results to the frontend as each node completes. Execution History records every run — inputs, outputs, per-node results, and timing — and lets you reload a past run's inputs onto the canvas for debugging or re-execution.
Team
Ceren Kaya Akgün
Founding Engineer
Ceren Kaya Akgün is a founding engineer at Heym with a background in distributed systems and AI/ML engineering. She led the design of Heym's visual workflow canvas and multi-agent orchestration engine, and has spent the last several years building production-grade AI systems at scale. She writes about the gap between AI research and real-world deployment — the tooling, tradeoffs, and architecture decisions that make the difference between a demo and a system that runs reliably in production.
Mehmet Burak Akgün
Founding Engineer
Mehmet Burak Akgün is a founding engineer at Heym with a background in backend engineering and platform infrastructure. He designed and built Heym's workflow execution engine, real-time streaming architecture, and Docker-based self-hosted deployment system. His work focuses on making powerful AI automation accessible to any team that wants full control over their data and infrastructure — no vendor lock-in, no data leaving your stack.