promptry 0.6.1

Regression protection for LLM pipelines

promptry

Sentry for prompts. Sentry catches when your code breaks. promptry catches when your prompts break — versions them, runs eval suites in CI, and flags regressions or drift against a baseline. Local-first. No SaaS.

from promptry import track, suite, assert_semantic

# track() content-hashes your prompt and stores a new version if it changed
prompt = track(system_prompt, "rag-qa")
response = llm.chat(system=prompt, ...)

# suites are regular Python functions. run them via CLI or in CI.
@suite("rag-regression")
def test_quality():
    response = my_pipeline("What is photosynthesis?")
    assert_semantic(response, "Converts light into chemical energy")

When a suite regresses against its baseline, promptry reports what changed:

Overall score: 0.910 -> 0.720  REGRESSION

Probable cause:
  -> Prompt changed (v3 -> v4)

Install

pip install promptry                       # core
pip install promptry[semantic]             # + semantic assertions (sentence-transformers)
pip install promptry[dashboard]            # + web dashboard
pip install promptry[semantic,dashboard]   # everything

Quick start

promptry init                              # scaffold project + starter eval
promptry run smoke-test --module evals     # run it

PASS test_basic_quality (142ms)
  semantic (0.891) ok

Overall: PASS  score: 0.891

Features

Feature	What it does
Prompt versioning	Content-hashed, automatic dedup
Eval suites	Semantic, schema, LLM-as-judge, JSON, regex, grounding assertions
Regression detection	Compare against baselines, get root cause hints
Drift detection	Catch slow quality degradation over time
Model comparison	Statistical comparison against historical baseline (not just snapshots)
Cost tracking	Token usage and cost per prompt, aggregated reports
Safety templates	25 starter jailbreak / injection / PII tests — add your own
MCP server	Expose everything as tools for Claude, Cursor, VS Code, etc.
Dashboard	Web UI for eval history, prompt diffs, model comparison, cost
JS/TS client	Ship prompt events from frontend/Node apps

Dashboard

pip install promptry[dashboard]
promptry dashboard

How it differs

	Promptfoo	DeepEval	RAGAS	LangSmith	promptry
Language	TypeScript	Python	Python	Python + JS	Python + JS
Local-first	Yes	Cloud push	Yes	SaaS only	SQLite
Prompt versioning	Via git + YAML	No	No	Prompt Hub	Automatic
Drift over time	No	No	No	Dashboards	Regression window
Root cause hints	No	No	No	No	Yes
Safety / red-team	Yes	Yes	No	No	25 starters
MCP server	Plugin	Partial	No	No	Native
Vendor	OpenAI-owned	Independent	Independent	LangChain	Independent
Cost	Free	Freemium	Free	Freemium	Free

Honest caveats: Promptfoo has more assertion types and a larger red-team corpus. RAGAS has the gold-standard RAG metrics (faithfulness, context precision, answer relevancy). LangSmith has better multi-user dashboards and deeper LangChain integration. promptry's niche is the combo of local SQLite + automatic versioning + CI-native + MCP server in one Python-first package.

GitHub Action

Run eval suites in CI with one line. View on Marketplace.

# .github/workflows/eval.yml
name: Eval
on: [push]
jobs:
  eval:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: bihanikeshav/promptry@v0.5.0
        with:
          suite: rag-regression
          module: evals
          compare: prod  # optional — compare against baseline

Input	Required	Default	Description
suite	Yes		Eval suite name
module	Yes		Python module containing the suite
compare	No		Baseline tag to compare against
python-version	No	3.12	Python version
extras	No	semantic	pip extras to install

MCP server

claude mcp add promptry -- promptry mcp    # Claude Code

Works with Claude Desktop, Cursor, Windsurf, VS Code. See full setup.

Documentation

The full guide covers all assertions, cost tracking, model comparison, safety templates, notifications, storage modes, JS client, CLI reference, MCP setup, and config options.

Honest caveats

• Early-stage. v0.6, solo-maintained, small user base. API is stable but bus-factor is one. Issues welcome.
• "No API keys" applies to the framework only. SQLite storage and the CLI need nothing. assert_llm, assert_grounded, and cost tracking all need your own LLM provider key.
• Drift detection is a rolling-window regression on scores. Works for steady degradation over a configurable window (default 30 runs). It is not a formal hypothesis test — see drift detection docs for exactly what it does and does not do.
• Safety templates are starters, not comprehensive coverage. 25 curated prompts across 6 categories. For serious red-teaming look at garak or PyRIT. Bring your own templates via templates.toml.
• Cost tracking uses hardcoded rate tables. Fine for rough estimates; won't reflect batching discounts, prompt caching, or provider price changes. Reconcile against invoices for finance.
• Auto-instrumentation is opt-in. promptry.integrations.openai and .litellm wrap clients automatically; otherwise you add track() manually. Explicit by default.
• No hosted multi-user UI. For that, look at LangSmith or Arize.

License

MIT