fzp

fzp (Fuzzy Processor)

Parallel LLM pipe filter. Reads text lines from stdin, sends each to an LLM in parallel, writes results to stdout preserving input order.

cat items.txt | fzp "Classify into: bug, feature, question"

Scope

fzp targets 1-shot batch tasks: classify, extract, translate, normalize. It deliberately stays small — multi-turn conversations, agent loops, and tool calling are out of scope. For those, reach for a Claude / OpenAI / Gemini SDK directly.

Install

cargo install fzp

Setup

fzp uses any OpenAI-compatible API. It's designed for lightweight, fast models.

fzp init

This creates ~/.config/fzp/config.toml with your API key, model, and endpoint.

To avoid storing the key in plaintext, replace api_key with api_key_command, whose stdout is used as the key:

api_key_command = "pass show fzp/openrouter"

api_key takes precedence when both are set.

Usage

# Inline prompt
<data> | fzp "Your prompt here"

# Named preset
<data> | fzp -p classify -v labels="bug,feature,question"

# Preset + extra instruction
<data> | fzp -p summarize "Respond in Japanese"

Options

Flag	Description	Default
-p NAME	Preset name	-
-v KEY=VALUE	Template variable (repeatable)	-
-m MODEL	Model override	from config.toml
-j N	Concurrency	64
--cache	Dedup identical input lines (skip duplicate API calls)	off
--list	List available presets	-

Built-in presets

Preset	Description	Variables
classify	Assign one label from a set	labels
summarize	One-sentence summary	-
translate	Translate text	lang
normalize	Extract fields as compact JSON	fields
filter	Output 1 (match) or 0 (no match)	condition

Examples

# Classify and count
cat items.txt | fzp -p classify -v labels="bug,feature,question" | sort | uniq -c

# Filter
paste items.txt <(cat items.txt | fzp -p filter -v condition="security-related") \
  | awk -F'\t' '$2 == "1"' | cut -f1

# Normalize to JSON
cat messages.txt | fzp -p normalize -v fields="name,topic,urgency"

Custom presets

Add presets to ~/.config/fzp/config.toml:

[prompt.my-preset]
template = "Your prompt with {{var}}"

Structured output

Attach a JSON Schema to constrain the model's output (provider-dependent enforcement; OpenRouter routes to providers that honor it):

[prompt.extract]
template = "Extract name and age."

[prompt.extract.output_schema]
type = "object"
required = ["name", "age"]
additionalProperties = false

[prompt.extract.output_schema.properties.name]
type = "string"

[prompt.extract.output_schema.properties.age]
type = "integer"

Claude Code plugin

fzp is also available as a Claude Code plugin. Add it to your project:

claude mcp add-plugin github.com/rail44/fzp

License

MIT