npm install -g @reqraft/cli

rp init

rp "add a button to export the report"

What it does

Reqraft is a local-first, open-source CLI that sits just before you send a prompt to Claude Code, Codex, OpenCode, ChatGPT, or another agent. It rewrites your request so the model receives:

• correct spelling and grammar,
• a clear structure,
• preserved technical terms, paths, and commands,
• local quality signals for known scope additions and disproportionate expansion.

Installation

npm install -g @reqraft/cli
# or
pnpm setup  # once, if pnpm has not configured its global bin directory yet
pnpm add -g @reqraft/cli

After pnpm setup, restart the terminal before running the global installation.

Two commands are provided and point to the same program:

rp "your request"
reprompt "your request"

Then run the guided setup and verify the resulting configuration:

rp init
rp auth login openai  # only when the selected provider key is not already in the environment
rp doctor
rp

rp init chooses the provider, model, profile, level and local preferences. Replace openai in the authentication command with the provider selected during setup. The final rp command opens the interactive interface.

To start the wizard again from its defaults, use rp init --reset.

Quick usage

Direct command

rp "fix the form on mobile"

Pipe input

echo "fix this without changing anything else" | rp

Read from file

rp --file request.md

Copy result to clipboard

rp "your request" --copy

Read from clipboard

rp --clipboard

Structured JSON output

rp "your request" --json

Starting with 0.2.0, JSON uses a locale-neutral, versioned envelope:

{
  "schemaVersion": 1,
  "ok": true,
  "result": {}
}

Failures use the same envelope with ok: false and a stable error code, optional params, and numeric exitCode. Human-readable translated messages are intentionally excluded so scripts receive the same JSON in every UI locale.

Diff view

rp "your request" --diff

Explain changes

rp "your request" --explain

Quality diagnostics

When a provider returns usable text, Reqraft keeps it available even if its best-effort fidelity checks identify an expansion or a known scope addition. Diagnostics are written to stderr and are available as structured data with --json.

rp "create a detailed architecture plan" --stats
rp "..." --max-output-tokens 2000
rp "..." --timeout 45000
rp "..." --fail-on-quality

See docs/reprompt-policy.md for the result, timeout, token-budget and fidelity contracts.

Profiles

Use --profile <name> to tune the rewriting style:

| Profile | Use case | | ------------ | ---------------------------------------- | | auto | The model picks the best-fitting profile | | clean | Grammar and light clarification | | code | Developer agents | | frontend | Frontend implementation | | web-design | Visual design and landing pages | | debug | Bugs and errors | | review | Code audits and reviews | | writing | Emails, messages, documents |

Local profiles

The table above lists the built-in profiles, which ship with Reqraft and are never modified. Alongside them you can keep as many local profiles as you like — one JSON file each, in Reqraft's profile directory. Both kinds work the same way with --profile <id> and both appear in the TUI picker, grouped by origin so you always know which is which.

rp profiles                       # list built-in and local profiles
rp profiles add                   # create one, guided
rp profiles add --file ./sav.reqraft-profile.json # import one, strictly validated
rp profiles edit sav              # change a local profile
rp profiles duplicate clean sav   # copy any profile into a new local one
rp profiles export sav            # print a portable JSON document
rp profiles remove sav            # delete a local profile, after confirmation

add asks for a name and suggests an id derived from it, which you can accept with Enter or replace. It then asks for a description, an optional built-in base to inherit from, the default level and the instructions.

Only local profiles can be edited or removed. A built-in profile is not a restriction to work around — duplicate gives you an editable copy of it:

rp profiles duplicate writing ton-maison --name "Ton maison"
rp profiles edit ton-maison
rp --profile ton-maison "réécris ce message"

Duplicating a built-in copies its instructions into the new file, so the copy stays standalone: it keeps working even if the built-in it came from later changes.

export writes the document to standard output, so notes and warnings go to standard error and a redirect captures JSON and nothing else:

rp profiles export ton-maison > ton-maison.reqraft-profile.json
rp profiles export ton-maison --output ./ton-maison.reqraft-profile.json   # or write it directly

Exporting a built-in renames it — clean becomes clean-copy — because the format refuses built-in ids. Without the rename the file would export cleanly and fail on import. Pass --as <id> to choose the id yourself:

rp profiles export clean --as ma-base > ma-base.reqraft-profile.json
rp profiles add --file ./ma-base.reqraft-profile.json

A local profile may inherit from one built-in profile through extends. The parent's instructions come first, yours follow, and your default level is kept:

{
  "schemaVersion": 1,
  "id": "sav",
  "name": "Support client",
  "description": "Reformule les réponses du support.",
  "extends": "clean",
  "defaultLevel": "standard",
  "instructions": "Réponds avec empathie, précision et une action claire."
}

Import is strict: an unknown field, a missing one or an id that collides with a built-in profile is refused with a message naming the problem, rather than being silently ignored.

Levels

rp --level minimal   # fix typos only
rp --level standard  # default: clarify and structure
rp --level complete  # rigorous brief, flag missing info

Providers

Reqraft supports multiple providers using native fetch:

• Anthropic
• OpenAI
• DeepSeek
• Mistral
• OpenAI Compatible (Ollama, LM Studio, gateways, etc.)

On macOS and Linux, the recommended option is the system credential manager:

rp auth login openai
rp auth status
rp auth logout openai

macOS uses Keychain and Linux uses Secret Service. Windows currently uses environment variables.

Environment variables remain available for CI, containers and unsupported credential stores:

export ANTHROPIC_API_KEY=...
export OPENAI_API_KEY=...
export DEEPSEEK_API_KEY=...
export MISTRAL_API_KEY=...

Environment variables take precedence over stored credentials.

Configuration

rp init                      # recommended first-run wizard
rp init --reset              # restart from defaults, with confirmation
rp config                    # show config
rp config get defaultProvider
rp config set defaultProvider openai
rp config path
rp config setup              # compatibility alias for rp init
rp config setup --reset      # compatibility alias for rp init --reset
rp doctor                    # verify config, keys and provider readiness

Language

Reqraft supports English and French for both the quick CLI and the interactive terminal UI. Automatic detection uses the terminal locale and falls back to English. Override it for one command or persist the preference:

rp --ui-locale fr "corrige ce texte"
rp --ui-locale en --help
rp config set uiLocale fr    # auto, en, or fr

For CI or temporary shell configuration, use REQRAFT_UI_LOCALE=en or REQRAFT_UI_LOCALE=fr.

The generated prompt language is independent from the interface language. auto preserves the input language; an explicit value asks the provider for a specific output language:

rp "rewrite this request" --output-language fr
rp config set outputLanguage auto

Changing uiLocale never changes the machine-readable JSON contract or the requested output language.

rp init never stores API keys in config.json. When a key is missing, it recommends rp auth login <provider> for secure system storage and also prints shell-specific environment-variable instructions.

Shell aliases

rp alias set p
rp alias set ask
rp alias list
rp alias remove p

Security

Reqraft detects common secrets locally before sending text to a provider:

• GitHub tokens
• OpenAI / Anthropic API keys
• AWS credentials
• Private keys
• Variables named SECRET, TOKEN, PASSWORD, or API_KEY

If a secret is detected, the CLI stops and suggests --redact-secrets or --force.

See SECURITY.md and docs/privacy.md.

Benchmark

pnpm benchmark
# or
pnpm exec tsx benchmark/runner.ts anthropic claude-haiku-4-5

Benchmarks are never run automatically because they consume API credits.

Contributing

See CONTRIBUTING.md and docs/development.md. Static analysis and coverage setup are documented in docs/code-quality.md. Release automation is documented in docs/development.md, and the UI/output language strategy in docs/i18n-feasibility.md.

License

MIT