28 Unix tools. One binary. Zero dependencies.·the missing coreutils for the agent era·try vrk in your browser at vrk.sh/play·curl vrk.sh/install.sh | sh - ready in 5 seconds·

The missing coreutils
for the agent era.

zero silent failures  •  composable by default  •  one static CLI binary

vrk grab https://api.example.com/docs \
  | vrk tok --check 8000 \
  | vrk prompt --system "extract the breaking changes" \
  | vrk validate --schema changes.json \
  | vrk kv set changelog
curl -fsSL vrk.sh/install.sh | sh

Or try it in your browser first. No install, no API key.

Try vrk in your browser

Pick a recipe, or type your own command. It runs on sample files, right in this page.

vrk, compiled to WebAssembly

# Strips the markdown, hides the API key, then checks that the result fits in 200 tokens.

Enter runs, Shift+Enter adds a line
    output 
    Press RUN. Nothing leaves your browser.

    files you can cat:

    22 of the tools run here. In real use vrk runs in your terminal and agent pipelines, on your own files. Install vrk.

    before

    import anthropic, tiktoken, json, re, jsonschema

client = anthropic.Anthropic()
with open("article.md") as f:
    text = f.read()

# Count tokens - abort if too long
enc = tiktoken.encoding_for_model("gpt-4")
if len(enc.encode(text)) > 8000:
    raise ValueError("over token budget")

# Redact secrets before sending
text = re.sub(r"sk-[a-zA-Z0-9]{48}", "[REDACTED]", text)

# Call the API
resp = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=1024,
    messages=[{"role": "user", "content": text}]
)
result = resp.content[0].text

# Validate the output
schema = json.load(open("schema.json"))
jsonschema.validate(json.loads(result), schema)

    after

    ~ vrk
    cat article.md \
  | vrk tok --check 8000 \
  | vrk mask \
  | vrk prompt --system 'Summarize' \
  | vrk validate --schema schema.json

    every vrk tool: stdin in, stdout out, exit 0/1/2

    over token limit?exit 1← pipeline stops
    bad JSON schema?exit 1← nothing downstream runs
    wrong flags?exit 2← usage error, never silent

    no more pipelines that look like they worked.

    batch jobs at scale

    ~ vrk
    find docs/ -name '*.md' -exec cat {} + \
  | vrk chunk --size 4000 \
  | vrk throttle --rate 60/m \
  | vrk prompt --field text --model gpt-4o --json

    chunked. paced. structured output. safe to rerun.

    PDFs and CSVs go in the same pipe

    Most of what you want an LLM to read is not markdown. Two tools turn it into text the next step can use, with no Python and no database server.

    vrk pdf: read a PDF as text →
    ~ vrk
    vrk pdf contract.pdf \
  | vrk chunk --size 4000 \
  | vrk prompt --field text \
      --system 'List every termination clause'

    Prints the text of each page, or JSONL with page numbers. A scanned PDF with no text exits 1, so the model never gets an empty page.

    vrk sql: query CSV files with SQL →
    ~ vrk
    vrk sql 'SELECT u.city, SUM(o.total) AS revenue
         FROM orders o JOIN users u ON u.id = o.user_id
         GROUP BY u.city' users.csv orders.csv \
  | vrk prompt --system 'Explain this revenue split'

    Each CSV or JSONL file becomes a table you can join and group. The result comes out as CSV or JSONL.

    Install

    brew tap vrksh/vrksh
brew install vrk
    go install github.com/vrksh/vrksh@latest
    curl -sSL https://vrk.sh/install.sh | sh

    Core

    vrk tok start here

    Count tokens. Gate pipelines before they fail.

    cat system-prompt.txt | vrk tok --check 8000
    vrk prompt

    LLM prompt - Anthropic/OpenAI, --schema, --retry, --explain

    cat article.md | vrk prompt --system 'Summarize the key findings in 3 bullet points'
    vrk chunk

    Token-aware text splitter - JSONL chunks within a token budget

    cat contract.pdf.txt | vrk chunk --size 4000 --overlap 200

    Pipeline

    vrk coax

    Retry wrapper - --times, --backoff, --on, --until

    vrk grab

    URL fetcher - clean markdown, plain text, or raw HTML

    vrk kv

    Key-value store - SQLite-backed, namespaces, TTL, atomic counters

    vrk mask

    Secret redactor - entropy + pattern-based, streaming

    vrk pdf

    PDF text extractor - Markdown, --text, JSONL per page

    vrk sql

    SQL over CSV or JSONL - SQLite in memory, CSV or JSONL out

    vrk sse

    SSE stream parser - text/event-stream to JSONL

    vrk throttle

    Rate limiter for pipes - --rate N/s or N/m

    vrk validate

    JSONL schema validator - --schema, --strict, --fix, --json

    + 16 utilities: jwt, epoch, uuid, and more →

    Lists every vrk tool to Claude Code and Cursor via MCP; the agent runs them in its shell. See agent integration →