FlowOS
Techmatic Systems
Visual AI workflow and agent platform: build on a canvas, edit any node's Python, and serve the same flow as REST, MCP or an OpenAI-compatible API.
- Integration families
- 111
- LLMs, vector stores, tools, observability
- Built-in components
- 523
- Test pass rate
- 100%
- Across 400 automated and manual cases
01
Problem
- Who
- Teams taking LLM apps and agents from prototype to production.
- Why it mattered
- Prototypes get rewritten as services, visual tools dead-end when needs outgrow the node palette, providers lock you in, and flows have no governance or rollback.
- Before
- A notebook demo, then a weeks-long rewrite whose logic drifts from what was validated.
02
Constraints
Technical
Ship as a Windows desktop application.
Infrastructure
Zero-touch setup: install Python and uv dependencies automatically on the user's machine.
Timeline
Delivered in the first three weeks, alongside AirDrive, a 1M-row dataset and the company website.
03
Architecture
- Component
- Key decision (click)
Author
- Visual canvas@xyflow/react, 33 starter templates
Run
- Graph execution engineFastAPI platform
- flowos-coreStateless, database-free executor
Serve
- REST API
- MCP serverProject-scoped
- OpenAI-compatible API
Govern
- Pluggable RBAC + audit log
- Flow versioning + rollback
- RS256/RS512 JWT, Fernet secrets
A flow built on the canvas is the deployable artifact: the same saved flow runs as a REST API, an MCP server, an OpenAI-compatible endpoint or an in-process Python call.
Governance (RBAC, audit, versioning) and production auth ship in the box; flowos-core runs flows without a database for serverless, CI and edge.
04
Key decisions
D1One artifact, four runtimes
- Alternatives considered
- Export code from the canvas and maintain it separately
- Why
- No rewrite between prototype and production, so logic can't drift.
D2Tool-calling-first design for the core platform
- Alternatives considered
- Cloud LLM (Claude API) vs. a local-LLM architecture
- Why
- Reduces AI inference dependency and operational cost.
05
Build highlights
- 01
Visual without a ceiling
Open any component's Python source and edit it in place; the canvas is a view over real code.
- 02
Governance that fails closed
RBAC on (subject, domain, object, action) with a queryable audit log; enabling the flag without a plugin can never widen visibility.
- 03
Versioned rollback
Snapshot, activate and roll back any flow with one API call.
- 04
Validated, not just demoed
400 automated and manual test cases at a 100% pass rate before release.
- 05
Zero-touch installer
Packaged with Electron, with an auto-installer that provisions Python and uv dependencies.
06
Results
100%
Pass rate
400 automated + manual test cases
400
Test cases
At release
How this was measured: Release test suite (automated and manual).
07
