Roadmap

​

Current status

Where pflow is today:

• Core workflow engine built on PocketFlow
• Node system — file, llm, http, shell, claude-code, and MCP bridge
• AI agent integration via CLI and MCP server
• Discovery — find nodes and workflows by describing what you need
• Template variables — connect node outputs to inputs
• Workflow validation with actionable error messages
• Execution traces for debugging
• Settings management — API keys, node filtering
• Unified model support — use any llm-supported provider for discovery and workflows
• Batch processing — process arrays of items through a single node (sequential or parallel)

​

Now

Getting pflow into users’ hands Current focus is preparing for public release:

• Completing user documentation
• Publishing to PyPI for easy installation
• Ensuring a smooth first-run experience

​

Next

Model discovery

• Show available models to agents based on configured API keys
• Help agents select appropriate models for different tasks

Proving the value

• Benchmark pflow’s efficiency using MCPMark evaluation
• Quantify token savings and latency improvements

​

Later

More expressive workflows Expanding what workflows can express:

• Conditional branching — if/else logic in workflows
• Task parallelism — run independent nodes concurrently (fan-out/fan-in)
• Nested workflow support

Better output control

• Structured output from LLM nodes (JSON schemas)
• Export workflows to standalone Python code
• Execution preview before running

Safer execution

• Sandbox runtime for shell commands
• Granular permission boundaries

Workflows as products

• Export workflows as self-hosted MCP server packages
• Share automation as installable tools

​

Vision

Long-term ideas on the radar:

• Discover and install MCP servers automatically
• Community registry for workflows and MCP servers
• Cloud execution for team use cases
• Workflows exposed as remote HTTP services

These are exploratory — they’ll take shape based on feedback and real-world usage. pflow improves through a direct feedback loop: agents build workflows, we identify where they struggle, we fix it. Because the surface area is finite — node types, template syntax, error messages — each fix compounds.

​

Get involved

---

Built by a developer who got tired of watching agents re-think the same tasks.
Questions or ideas? Reach out — andreas@pflow.run