Roadmap

Design Principles

These are not aspirational. They are constraints that shape every decision.

Compiled agent. AI generates code at creation time. Deterministic code runs at execution time. No LLM in the hot path. Zero runtime token cost.

Local-first. Runs on your laptop with no account, no cloud dependency, no telemetry. Cloud deployment is opt-in.

You own the code. Every generated file — YAML, SQL, scripts — lives in your Git repository. Fork it. Edit it. Delete Radhflow and keep your code.

Contract-first. node-spec.yaml defines inputs, outputs, and params before code exists. Schemas are checked at construction time, not at runtime.

Docker boundary. One container image. Same image for local, SaaS, enterprise. Docker is the only runtime boundary — no ambient system dependencies.

EU sovereignty. Hosted infrastructure runs on Hetzner and Scaleway. No US hyperscaler in the data path. Customer data stays in EU jurisdiction.

No customer code in cloud. The hosted service executes pipelines but does not retain generated code. Pipeline artifacts go to customer-owned storage.

v0.1 — Local Engine MVP

Core engine. CLI. File I/O. Data operations.

rf init / rf run / rf validate / rf inspect CLI commands
Pipeline YAML parser and graph executor
Four data types: Value, Record, Table, Stream
NDJSON interchange with .schema.json companions
DuckDB-backed data operations: filter, map, sort, limit, dedup, join, group, sql
File connectors: CSV, JSON, NDJSON read/write
Schema validation at construction time
Topological execution order with error propagation
Git-backed workspace with auto-commit

v0.2 — CLI Nodes and Connectors

Shell tools. Platform nodes. External data sources.

nix-shell integration for reproducible CLI environments
bubblewrap sandboxing for CLI nodes (isolated filesystem, no network by default)
Platform nodes: curated CLI tools with pre-approved sandbox configs
HTTP / REST API connector with auth, pagination, rate limiting
Google Sheets connector (read and write)
Browser data extraction via Playwright
Parallel node execution with configurable chunking
Router nodes for conditional branching
Secret management in local credential vault

v0.3 — Visual Canvas

React Flow editor. Visualizations. MCP server.

React Flow canvas: drag nodes, draw edges, see data flow
Live data preview on edges (sample rows)
YAML and canvas stay in sync — edit either
Pipeline visualization: execution status, timing, row counts
@radh/flow-mcp server: create, validate, run pipelines from AI agents
MCP integration with Claude Code, Cursor, Copilot
rf serve for local web UI
Pipeline templates and quickstart scaffolding

v0.4 — Hosted Service

EU infrastructure. Scheduling. Credentials. Payments.

Hetzner / Scaleway deployment (EU-only)
Cron-based pipeline scheduling
Managed credential vault with encryption at rest
Customer-scoped S3 storage for pipeline artifacts
Team workspaces with role-based access
Pipeline execution logs and monitoring
x402 payment integration for metered billing
API key management for external integrations
Webhook triggers for event-driven pipelines

v0.5 — Isolation and Registry

Firecracker microVMs. Community nodes. GPU.

Firecracker microVM isolation for agent-generated code
~125ms boot time, ephemeral filesystem, no state leakage
License whitelist enforcement: green (MIT, BSD, Apache), amber (GPL subprocess), red (AGPL blocked)
Community node registry: publish, discover, install third-party nodes
Node versioning with semantic compatibility checks
GPU node support for ML inference workloads
Pipeline composition: nest pipelines as nodes in other pipelines
Audit log for hosted execution (who ran what, when, with what data)

Status

v0.1 is in active development. This roadmap reflects current plans and will evolve. Each version builds on the previous — no version ships without its predecessors being stable.