Thanks to all of our amazing contributors.
Prompts to agents, with observability
Prompty is a markdown file format for LLM prompts — define model config,
inputs, tools, templates, and agentic controls in YAML frontmatter, then
execute across Python, TypeScript, C#, and Rust.
One file, complete prompt
A .prompty file pairs YAML frontmatter with a markdown prompt body
in a single, portable asset. Define your model, inputs, tools, and template —
then execute it from VS Code or any Prompty runtime with one command.
- Model configuration and connection details
- Input schema with types and defaults
- Tool definitions (function, MCP, OpenAPI)
- Template rendering with Jinja2 or Mustache
Author in VS Code
The Prompty VS Code extension is your prompt engineering workbench. Write prompts with full language support, run them with F5, and inspect every detail in the built-in trace viewer.
- Syntax highlighting and autocomplete
- Connection management for OpenAI, Anthropic, and Microsoft Foundry
- Live preview and interactive chat mode
- One-click execution with inline results
Trace everything
Every execution produces a detailed trace — messages sent, tokens used, latency, and the raw API response. Debug prompt issues in seconds, not hours.
- Built-in console, JSON file, and OpenTelemetry backends
- VS Code trace viewer with conversation, timing, and raw views
- Decorate your own functions with
@tracefor end-to-end visibility - Zero-overhead when no tracer is registered
The pipeline
Prompty processes every prompt through a four-stage pipeline, then adds
turn() when you need an agentic loop around that same prompt asset.
- Render — expand template variables with Jinja2 or Mustache
- Parse — split role markers into a message list
- Execute — call the LLM via any registered provider
- Process — extract content, parse structured output, stream chunks
1
Portable by design
Same .prompty file works across VS Code and the Prompty runtimes — no rewriting.
2
Any provider
OpenAI, Microsoft Foundry, Anthropic — or build your own executor in minutes.
3
Observable by default
Built-in tracing with pluggable backends — console, files, or OpenTelemetry.
Agentic runtime, same prompt file
Prompty v2 adds a shared agentic turn runtime on top of the normal
render → parse → execute → process pipeline. Define tools in frontmatter,
provide handlers from host code, and call turn() when the model should
use tools or carry a governed conversation forward.
- Function tools with local callables
- MCP, OpenAPI, and custom tool declarations
- Runtime controls for guardrails, steering, cancellation, and retries
- Context budgets, compaction, and memory-aware conversation turns
- Event streams and traces for every model/tool loop iteration
Open source, open standard
Prompty is built on the premise that even with increasing complexity in AI, a fundamental unit remains prompts. The format is open, extensible, and designed to grow with the ecosystem.
- TypeSpec-defined schema with generated model code
- Entry-point-based plugin system for custom providers
- MIT licensed — contributions welcome