Run LLM .prompt files from your shell with a single-file Python script.

Dotprompt is an prompt template format for LLMs where a .prompt file contains the prompt and metadata (model, schema, config) in a single file. You can use it to run LLM prompts and get structured responses right in your shell.

Quick start | Examples | Tools | Template syntax | Configuration | Providers | Caching | Spec compliance

curl -O https://raw.githubusercontent.com/chr15m/runprompt/main/runprompt
chmod +x runprompt

Create hello.prompt:

---
model: anthropic/claude-sonnet-4-20250514
---
Say hello to {{name}}!

Run it:

export ANTHROPIC_API_KEY="your-key"
./runprompt hello.prompt '{"name": "World"}'
# alternatively pass data via STDIN.
echo '{"name": "World"}' | ./runprompt hello.prompt

(You can get an Anthropic key from here: https://console.anthropic.com/settings/keys see below for other keys.)

In addition to the following, see the tests folder for more example .prompt files.

---
model: anthropic/claude-sonnet-4-20250514
---
Summarize this text: {{STDIN}}

cat article.txt | ./runprompt summarize.prompt

The special {{STDIN}} variable always contains the raw stdin as a string. Both {{STDIN}} and {{ARGS}} are always available; if input is JSON it's also parsed into individual variables.

Pass arguments directly on the command line:

---
model: anthropic/claude-sonnet-4-20250514
---
Process this: {{ARGS}}

./runprompt process.prompt Hello world, please summarize this text.

The special {{ARGS}} variable contains all arguments after the prompt file, joined with spaces. The {{INPUT}} variable contains STDIN if provided, otherwise ARGS.

Extract structured data using an output schema:

---
model: anthropic/claude-sonnet-4-20250514
input:
  schema:
    text: string
output:
  format: json
  schema:
    name?: string, the person's name
    age?: number, the person's age
    occupation?: string, the person's job
---
Extract info from: {{text}}

echo "John is a 30 year old teacher" | ./runprompt extract.prompt
# {"name": "John", "age": 30, "occupation": "teacher"}

Schema uses Picoschema format. Fields ending with ? are optional. The format is field: type, description.

Pipe structured output between prompts:

echo "John is 30" | ./runprompt extract.prompt | ./runprompt generate-bio.prompt

The JSON output from the first prompt becomes template variables in the second.

Make .prompt files directly executable with a shebang:

#!/usr/bin/env runprompt
---
model: anthropic/claude-sonnet-4-20250514
---
Hello, I'm {{name}}!

chmod +x hello.prompt
echo '{"name": "World"}' | ./hello.prompt

Note: runprompt must be in your PATH, or use a relative/absolute path in the shebang (e.g. #!/usr/bin/env ./runprompt).

Override frontmatter values from the command line:

./runprompt --model anthropic/claude-haiku-4-20250514 hello.prompt
./runprompt --output.format json extract.prompt

Use dot notation for nested values (e.g. --output.format). Pass template variables via stdin or as JSON argument:

echo '{"name": "Alice"}' | ./runprompt hello.prompt
./runprompt hello.prompt '{"name": "Alice"}'

See --help for all options.

Attach local files (supports globs) and URLs to the prompt context.

From frontmatter:

---
model: anthropic/claude-sonnet-4-20250514
files:
  - README.md
  - src/**/*.py
---
Review these files and suggest improvements.

From the CLI (also supports globs). --file is an alias for --read:

./runprompt --read README.md --read "src/**/*.py" review.prompt
./runprompt --file README.md review.prompt

You can combine both: frontmatter files: entries plus any --read/--file flags are all included.

When attachments are loaded, runprompt prints a yellow message to stderr for each loaded file/URL.

Tools allow the LLM to call Python functions during prompt execution. Define tools as Python functions with docstrings, and the LLM can use them to perform actions like reading files, making API calls, or interacting with the system.

Create a Python file with functions. Any function with a docstring becomes a tool:

# my_tools.py
def get_weather(city: str):
    """Gets the current weather for a city.
    
    Returns the temperature and conditions for the specified location.
    """
    # Your implementation here
    return {"temp": 72, "conditions": "sunny"}

def calculate(expression: str):
    """Evaluates a mathematical expression.
    
    Use this for arithmetic calculations.
    """
    return eval(expression)

Reference tools in the frontmatter using Python import syntax:

---
model: anthropic/claude-sonnet-4-20250514
tools:
  - my_tools.*
---
What's the weather in Tokyo and what's 15% of 847?

Tool specifications:

• module.* - Import all functions with docstrings from module.py
• module.function_name - Import a specific function

When the LLM wants to call a tool, you'll be prompted to confirm:

$ ./runprompt weather.prompt

I'll check the weather for you.

Tool call: get_weather
Arguments: {"city": "Tokyo"}

Run this tool? [y/n]: y

The weather in Tokyo is currently 72°F and sunny.

Mark tools as "safe" to allow them to run without confirmation when using --safe-yes, env var RUNPROMPT_SAFE_YES, or config safe_yes:

# my_tools.py
def get_weather(city: str):
    """Gets the current weather for a city (read-only operation)."""
    return {"temp": 72, "conditions": "sunny"}

get_weather.safe = True  # Mark as safe

def delete_file(path: str):
    """Deletes a file from the filesystem."""
    os.remove(path)
    return "deleted"

# Not marked as safe - will always prompt for confirmation

Run with --safe-yes or env var RUNPROMPT_SAFE_YES or config var safe_yes to auto-approve safe tools:

./runprompt --safe-yes weather.prompt

With --safe-yes:

• Tools marked with fn.safe = True run without prompting
• Tools without the safe attribute still prompt for confirmation

This is useful for automation where you trust certain read-only or low-risk operations but still want confirmation for potentially dangerous actions.

Tools are searched for in:

1. The current working directory
2. The directory containing the prompt file
3. Additional paths via --tool-path or config file
4. Default config tool directories (if they exist):
   • ./.runprompt/tools
   • $XDG_CONFIG_HOME/runprompt/tools (default: ~/.config/runprompt/tools)
   • ~/.runprompt/tools

./runprompt --tool-path ./my_tools --tool-path /shared/tools prompt.prompt

All paths are searched in addition to any tool_path entries in config files.

Type hints on function parameters map to JSON Schema types:

Parameters without type hints default to string.

If a tool raises an exception, the error is sent back to the LLM which can decide how to proceed:

Tool call: read_file
Arguments: {"path": "missing.txt"}

Run this tool? [y/n]: y
FileNotFoundError: [Errno 2] No such file or directory: 'missing.txt'

I couldn't read that file because it doesn't exist. Would you like me to try a different path?

Runprompt includes builtin tools that can be used without creating external Python files:

---
model: anthropic/claude-sonnet-4-20250514
tools:
  - builtin.fetch_clean
---
Please summarize this page: {{ARGS}}

Available builtin tools:

The calculator, datetime, fetch_clean, sleep, and ask_user tools are marked as safe (auto-approved with --safe-yes). The shell and write_file tools require confirmation since they can run commands / modify the filesystem.

Some builtin tools are "factories" that take arguments to create a specialized tool:

---
model: anthropic/claude-sonnet-4-20250514
tools:
  - builtin.write_file('output.txt')
---
Write a haiku about coding to the file.

The write_file('output.txt') creates a tool that can only write to output.txt. The LLM provides the content, but cannot choose the path - this is a security feature that limits what files the LLM can modify. The created tool is exposed to the LLM with a filename-specific name (slug), e.g. write_file_output_txt.

Use builtin.* to import all builtin tools, or builtin.tool_name for a specific one.

Files or functions starting with _ are excluded from wildcard imports:

# _helpers.py - this entire file is excluded from wildcard imports
# my_tools.py
def _private_helper():  # excluded from wildcard imports
    """Internal helper function."""
    pass

def public_tool():  # included in wildcard imports
    """A tool available to the LLM."""
    pass

Use _ prefix for helper functions you don't want exposed as tools.

Templates use a useful subset of Handlebars/Mustache syntax. Supported features:

• Variable interpolation: {{variableName}}, {{object.property}}
• Comments: {{! this is a comment }}
• Conditionals: {{#if key}}...{{/if}}, {{#if key}}...{{else}}...{{/if}}
• Negated conditionals: {{#unless key}}...{{/unless}}, {{#unless key}}...{{else}}...{{/unless}}
• Iteration: {{#each items}}...{{/each}} with @index, @first, @last, @key
• Sections: {{#key}}...{{/key}} (renders if truthy)
• Inverted sections: {{^key}}...{{/key}} (renders if falsy)

Falsy values: false, 0, "" (empty string), [] (empty list), missing/undefined variables.

Configuration values can be set from config file, env var or command line flag, with flags overriding env vars which override config file settings.

1. Config files (lowest priority, loaded in order):
   • ~/.runprompt/config.yml
   • $XDG_CONFIG_HOME/runprompt/config.yml (default: ~/.config/runprompt/config.yml)
   • ./.runprompt/config.yml (project-local)
2. Environment variables (RUNPROMPT_* prefix)
3. CLI flags (highest priority)

API keys can be set via config file, RUNPROMPT_* env var, or native env var:

Priority for API keys: config file, env var, then flag as fallback.

# ./.runprompt/config.yml, ~/.config/runprompt/config.yml, or ~/.runprompt/config.yml
model: openai/gpt-4o
default_model: anthropic/claude-sonnet-4-20250514  # fallback if model not set anywhere
cache: true
safe_yes: true
tool_path:
  - ./tools
  - /shared/tools
openai_api_key: sk-...

The default_model is used as a fallback when no model is specified in the prompt file, config, environment, or CLI. This lets you set a preferred model that's used only when nothing else specifies one.

Use base_url to point at any OpenAI-compatible endpoint:

# Via config file
# base_url: http://localhost:11434/v1

# Via environment variable
export RUNPROMPT_BASE_URL="http://localhost:11434/v1"

# Via CLI flag
./runprompt --base-url http://localhost:11434/v1 hello.prompt

# Legacy env vars also work (checked in this order)
export OPENAI_BASE_URL="http://localhost:11434/v1"
export OPENAI_API_BASE="http://localhost:11434/v1"  # OpenAI SDK v0.x style
export BASE_URL="http://localhost:11434/v1"

When a custom base URL is set, the provider prefix in the model string is ignored and the OpenAI-compatible API format is used.

Use -v or set verbose: true to see request/response details:

./runprompt -v hello.prompt

Models are specified as provider/model-name:

OpenRouter provides access to models from many providers (Anthropic, Google, Meta, etc.) through a single API key.

Enable response caching to avoid redundant API calls during development:

# Enable caching with -c or --cache
./runprompt --cache hello.prompt

# Second run with same input uses cached response
./runprompt --cache hello.prompt

You can also enable the cache across a whole pipeline with the env var:

export RUNPROMPT_CACHE=1; echo "..." | ./runprompt a.prompt | ./runprompt b.prompt

Cached responses are stored in ~/.cache/runprompt/ (or $XDG_CACHE_HOME/runprompt/), based on the inputs applied to the template and frontmatter.

See --help for more information.

This is a minimal implementation of the Dotprompt specification. Not yet supported:

• Multi-message prompts ({{role}}, {{history}})
• Helpers ({{json}}, {{media}}, {{section}})
• Model config (temperature, maxOutputTokens, etc.)
• Partials ({{>partialName}})
• Nested Picoschema (objects, arrays of objects, enums)

The YAML parser is a minimal implementation that handles only the subset of YAML needed for Dotprompt frontmatter (simple key-value pairs, nested objects, and lists). It will likely fail on complex YAML features like anchors, multi-line strings, or flow syntax. If PyYAML is installed (pip install pyyaml), it will be used instead for full YAML support.