Open-source prompt infrastructure
AI features start with a few prompt strings. Then model settings, tools, provider quirks, context limits, environment overrides, tests, and customer-specific behavior start spreading across the codebase. PromptOpsKit turns that prompt glue into versioned assets that live in Git and ship with your app.
No hosted dashboard. No gateway required. No vendor lock-in. Keep your SDK, auth, routing, observability, and billing.
npm install promptopskit
npx promptopskit init
npx promptopskit skillArchitecture at a glance
Author prompt assets in Markdown, compose shared standards, validate in CI/CD, compile for production, and render provider-ready API bodies.
PromptOpsKit is for developers and tech leads shipping AI features in real applications: SaaS products, copilots, internal tools, support workflows, developer tools, and agentic product experiences.
Before
After
includes and defaults.md avoid copy-paste driftenvironments and tiers handle overrides cleanly.test.yaml sidecars keep deterministic test behaviorSee it working
A walkthrough of how PromptOpsKit structures prompts and renders provider-ready request payloads in a normal codebase.
Start with a readable Markdown asset, then grow into the full schema as needed. View full schema docs.
---
id: support/reply
provider: openai
model: gpt-5.4
includes:
- ./shared/tone.md
context:
inputs:
- name: user_message
non_empty: true
reject_secrets: true
environments:
dev:
model: gpt-5.4-mini
---
# System instructions
You are a helpful support assistant.
# Prompt template
{{ user_message }}PromptOpsKit renders provider-shaped request payloads. Your app keeps SDK choice, auth, retries, routing, observability, and billing.
import { createPromptOpsKit } from 'promptopskit';
const kit = createPromptOpsKit({ sourceDir: './prompts' });
const { request } = await kit.renderPrompt({
path: 'support/reply',
provider: 'openai',
environment: 'prod',
variables: {
user_message: 'How do I reset my password?'
}
});
// request.body is provider-ready
await fetch('https://api.openai.com/v1/chat/completions', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
Authorization: `Bearer ${process.env.OPENAI_API_KEY}`
},
body: JSON.stringify(request.body)
});Keep compression opt-in and close to the prompt. Use backend compression, conservative local extraction, JSON-to-TOON preprocessing, or code compaction, then read a lightweight operation-level token-savings summary from the render result.
compression:
heuristic:
enabled: true
mode: conservative
query_variable: user_question
json_to_toon: true
code:
enabled: false
context:
inputs:
- name: account_context
compression:
heuristic:
enabled: true
mode: conservative
- name: source_code
compression: code
# Prompt template
Context: {{ account_context | compress }}
Payload: {{ json_payload | toon }}
Source: {{ source_code | compact }}The local token-compression approach credits Jason Kneen's open-thetokenco. TOON preprocessing is inspired by the MIT-licensed TOON project by Johann Schopplich. Invalid TOON inputs are preserved with a warning, and code compaction skips backend text compression.
Testing for developers and teams
Sidecar .test.yaml fixtures let local development, unit tests,
and CI run deterministic prompt checks without provider calls.
prompts/
├── hello.md
└── hello.test.yaml
tests/
└── hello.prompt.test.mjscases:
- name: basic-greeting
variables:
name: "World"
response:
message: "Hello, World! How can I help you today?"Validate prompt behavior and expected outputs before merge without flaky network dependencies.
Develop UI and downstream logic against deterministic model-shaped responses.
Review prompt text, model settings, test cases, and expected outputs in one PR.
Put tone, policy, and safety guidance in shared files and include them across prompts. Update once, apply everywhere.
Define common provider and model settings per folder, then override per prompt only when needed.
Keep dev/prod and free/pro differences in one prompt asset with explicit, reviewable overrides.
Keep required fields, size limits, and allow/deny checks close to the prompt template so behavior is explicit and testable.
Render OpenAI Chat, OpenAI Responses, Anthropic, Gemini, OpenRouter, and LLMAsAService request bodies while keeping your app's auth, retries, observability, billing, and routing.
PromptOpsKit is the repo-native layer between prompt strings and production AI calls. Use it when you have outgrown hardcoded prompts but do not want a hosted prompt-management platform in your path.
When AI features get serious
PromptOpsKit handles the open-source prompt asset layer. As usage grows, teams often need adjacent controls for provider cost, routing, customer attribution, entitlements, and billing.
When you need routing, caching, cost controls, customer attribution, and gateway-level reliability, pair PromptOpsKit with LLMAsAService.
When AI features need metering, entitlements, limits, alerts, or usage-based billing, pair PromptOpsKit with UsageTap.
How it compares
Use GitHub Models for playgrounds, model comparison, and evals. Use PromptOpsKit when prompt behavior needs to live in your runtime with validated inputs, composition, overrides, compiled artifacts, and provider-ready request bodies.
Fits alongside
PromptOpsKit does not replace eval, tracing, or orchestration tools. It focuses on repo-native prompt assets and runtime request rendering.
Store prompt behavior in reviewable Markdown with YAML front matter.
Reuse shared instructions and baseline settings without duplication.
Keep dev/prod and plan-specific behavior in one prompt source.
Define required values, size limits, and content checks close to prompt text.
Opt into TheTokenCompany, conservative local extraction, TOON preprocessing, or code compaction with token-savings summaries.
Render provider-ready bodies for OpenAI, Anthropic, Gemini, OpenRouter, and LLMAsAService.
Test behavior deterministically in local dev and CI without model calls.
Neither. It is a repo-native runtime layer that structures prompt behavior and renders request bodies while your app keeps transport.
Yes. Use .test.yaml sidecars and PromptOpsKit testing helpers for
deterministic local and CI workflows.
Yes. PromptOpsKit returns request payloads only, so auth, retries, routing, observability, and billing stay in your stack.
PromptOpsKit manages prompt assets. Use LLMAsAService for provider routing, caching, and cost controls. Use UsageTap for customer usage, entitlements, limits, and usage-based billing.
Start with Getting Started, then review schema and testing docs for advanced patterns.
MIT-licensed. Runs in your codebase. Validates in CI. Compiles for production. No hosted service required.
View repository