v0.2 · early alpha · MCP record/replay in dev
Specs in Git. Signed releases. Test harnesses for JUnit, pytest, and Vitest. No new SaaS, no new auth — promptLM lives in the Git infra your team already trusts.
03 — capabilities
PROMPTS AS SPEC
Every prompt is a spec file — vendor, model, parameters, messages. Versioned in Git. Released as a signed JAR your app pins.
# promptlm.yaml id: customer_support group: support version: 1.0.0 request: vendor: openai model: gpt-4o parameters: { temperature: 0.6, maxTokens: 256 } messages: - role: system content: You are a support assistant. - role: user content: "Summarize: {{ticket}}"
STUDIO CATALOG
Browse, search, and filter the inventory. See status, executions, and latency at a glance. Syncs from Git on push.
IN YOUR APP
Released prompts ship as a JAR your app depends on. Load by id, fill placeholders, send. Vendor, model, and parameters travel with the spec — your code never drifts from the prompt.
// SupportService.java Prompt prompt = PromptLoader.load("customer_support"); // pinned v1.0.0 in pom.xml String filled = prompt.fill(Map.of("ticket", ticket)); // vendor, model, parameters come from the spec — not your code String reply = llm.send(filled, prompt.model(), prompt.parameters());
TEST SUPPORT
Each released prompt ships with golden responses captured during dev. Your tests replay them locally via WireMock — no real LLM calls, no tokens billed, deterministic in CI.
@InjectResponse for equality assertions
// SupportServiceTest.java @EnablePromptWireMock // boots WireMock on a local port, class SupportServiceTest { // pre-stubbed from your prompt repo @Test void summarizesTicket(@InjectResponse(id = "customer_support") String recorded) { String reply = supportService.summarize(ticket); // your prod code POSTs to the LLM // The LLM URL points at WireMock, so the recorded reply is returned. // No real model call, no tokens billed, deterministic across runs. assertThat(reply).isEqualTo(recorded); } }
# test_support_service.py from promptlm.testing import enable_prompt_mock, inject_response @enable_prompt_mock # boots a mock LLM on a local port, @inject_response("customer_support") # pre-stubbed from your prompt repo def test_summarizes_ticket(recorded: str): reply = support_service.summarize(ticket) # your prod code POSTs to the LLM # The LLM URL points at the mock, so the recorded reply is returned. # No real model call, no tokens billed, deterministic across runs. assert reply == recorded
// supportService.test.ts import { enablePromptMock, injectResponse } from '@promptlm/testing'; enablePromptMock(); // boots a mock LLM on a local port, test('summarizes ticket', () => { // pre-stubbed from your prompt repo const recorded = injectResponse('customer_support'); const reply = supportService.summarize(ticket); // your prod code POSTs to the LLM // The LLM URL points at the mock, so the recorded reply is returned. // No real model call, no tokens billed, deterministic across runs. expect(reply).toBe(recorded); });
COMING SOON ROADMAP
Capture every tool call against your real MCP servers, freeze the bytes, and replay them deterministically in CI. No flaky network, no vendor rate limits, no surprise charges. Same story as prompt replay — extended to the tools your agent actually uses.
GET STARTED