folpipe 0.1.0

Folder-based agentic AI pipeline framework

pipelinex

Folder-based agentic AI pipeline framework. No programming required to use.

pip install folpipe
folpipe run ./my-pipeline --watch

---

The idea

A pipeline is a folder. Instructions live in plain-English markdown files. Tools are drop-in folders. The model is swappable with one line.

my-pipeline/
├── pipeline.yaml       ← steps, model, routing
├── SKILL.md            ← global instructions (plain English)
├── .env                ← API keys
├── input/              ← drop files here
├── output/             ← results land here
├── tools/              ← drop-in custom tools
├── step-01-ingest/
│   └── SKILL.md        ← what this step does
└── step-02-process/
    ├── SKILL.md
    └── sub-01-chunk/   ← sub-steps, recursively
        └── SKILL.md

Zip the folder, share it, version it in git. No server, no database, no cloud account.

---

Quick start

1. Install

pip install pipelinex

2. Create a pipeline

pipelinex new my-pipeline
cd my-pipeline

3. Configure the model — edit .env:

DEEPSEEK_API_KEY=sk-...

4. Write your step — edit step-01-start/SKILL.md:

# Summarise input

Read the file from input/ and write a 3-paragraph summary to output/summary.md.

## Task

Use read_file to read the input file.
Use write_file to write output/summary.md.
Leave a handoff note when done.

5. Run it

cp myfile.txt input/
pipelinex run . --watch

---

What it can do

Linear pipelines — steps run in order, each hands off to the next via shared state.

Agent branching — the model decides which step comes next from a declared whitelist:

- id: step-03-validate
  can_goto: [step-04-output, step-05-partial-output]

Parallel dispatch — the model fires multiple dispatch_task calls in one response; the runner detects the batch and executes them concurrently:

## Task
Process all documents at once — send them all for chunking
in the same breath rather than one at a time.

Sub-steps — steps contain named sub-step folders with their own SKILL.md and tools:

dispatch_task(task="chunk intro-to-ml.txt", substep="sub-01-chunk", context={...})

Human-in-the-loop — steps can pause for console input, file review, or a custom tool (Slack, email, webhook):

- id: step-03-review
  human_input:
    mode: console
    prompt: "Approve this output? (yes/no):"

Model-driven context — each step's ## Context section tells the model what state to pay attention to and what to ignore. Skip signals are respected; the model never sees irrelevant data.

Any model — DeepSeek, Anthropic, OpenAI, Ollama, Groq, or any OpenAI-compatible endpoint. Change one line in pipeline.yaml. Optional per-step overrides for cost optimisation.

---

Custom tools

Drop a folder into tools/:

tools/
└── send_slack/
    ├── tool.json    ← MCP-compatible schema + deps declaration
    └── run.py       ← stdin JSON in, stdout JSON out

{
  "name": "send_slack",
  "description": "Post a message to a Slack channel. Use for approvals and failures that need human eyes.",
  "inputSchema": { ... },
  "deps": ["slack-sdk"]
}

Dependencies declared in deps are installed automatically before the first run. No manual pip install.

Any language works — run.py, run.sh, or run.js.

---

Example pipelines

doc-pipeline

Ingests documents → chunks → embeds → indexed output.

pipelinex run ./doc-pipeline --watch

Demonstrates: parallel dispatch, sub-steps, agent branching, partial-failure routing.

cv-pipeline

Job URL → developer profile → tailored CV + styled PDF.

pipelinex run ./cv-pipeline --input "https://example.com/jobs/engineer" --watch

Demonstrates: custom tools (fetch_page, render_pdf), linear pipeline, HTML→PDF via Edge headless.

---

CLI

pipelinex run ./my-pipeline               # run
pipelinex run ./my-pipeline --watch       # run with live output
pipelinex run ./my-pipeline --from step-03  # resume after failure
pipelinex run ./my-pipeline --dry-run     # validate config without running
pipelinex new my-pipeline                 # scaffold new pipeline
pipelinex new step step-05-review --in ./my-pipeline
pipelinex new tool send_email --in ./my-pipeline/tools
pipelinex validate ./my-pipeline          # check config
pipelinex tools list --pipeline ./my-pipeline
pipelinex tools install ./my-pipeline     # pre-install tool deps
pipelinex log ./my-pipeline               # show run log
pipelinex log ./my-pipeline --errors      # show last error report

---

Documentation

• Getting Started
• Pipeline Structure — pipeline.yaml and SKILL.md reference
• State & Handoffs
• Tools — built-ins, custom tools, resolution order
• Execution Patterns — linear, branching, loops, dispatch, sub-steps
• Context Management
• Human Input
• CLI Reference
• Examples

---

Requirements

• Python 3.10+
• A model with tool calling support (DeepSeek, Claude, GPT-4, Llama 3.1+)
• For PDF rendering in cv-pipeline: Edge or Chrome installed