Pyrula

Every step, model call and tool call is journaled, so a run resumes after a crash, reruns with a change, and replays as a test.

pip install "pyrula[agents]"
pyrula init
pyrula dev app.py

The scaffold talks to a scripted model, so it needs no key, broker or account. Until release, install is a source build (Install).

Run, what-if, case, CI

	On your machine	With Pyrula Cloud
Run	pyrula dev journals every step, model and tool call. A completed step is never re-run after a crash.	Every worker’s runs in one timeline (Connect).
What-if	pyrula case run --prompt-file or --model: model calls run live, recorded tool results answer the rest.	Fork a finished run and change its input, a step result, the model or the prompt. Writes denied, budget on (What-if runs).
Case	pyrula dev keeps the newest finished run in .pyrula/last/.	Save a run or a what-if as a case, stored in your fixture store (Saved cases).
CI	pytest --pyrula-cases cases: no model call, no network.	pyrula ci run or the replay-check GitHub Action (Replay checks in CI).

Record once, iterate free walks the left column.

A safe agent

from pyrula.agents import AgentContext, AgentLimits, ToolContext, agent, tool


@tool(read_only=True)
async def lookup_order(ctx: ToolContext, order_id: str) -> dict:
    return {"status": "delivered", "total_cents": 25_000}


@tool(requires_approval=lambda args: args["cents"] > 10_000, approval_timeout=3600)
async def refund(ctx: ToolContext, order_id: str, cents: int) -> str:
    return f"refunded {cents} cents on {order_id}"


@agent(limits=AgentLimits(max_cost_usd=2.50, max_total_tokens=200_000))
async def support(ctx: AgentContext, message: str) -> str:
    return await ctx.llm.run([{"role": "user", "content": message}], tools=[lookup_order, refund])

• A refund over $100 parks the run until a person answers from the dashboard or the resume endpoint; an hour with no answer denies it (Approval). pyrula dev (embedded mode) refuses approval_timeout, so run a worker for this agent. Pydantic AI tools get no approval. examples/tool_approval.py runs an approval and a rejection.
• The caps are checked before each live model call, so a run can overshoot by one in-flight call per concurrent caller (Spend caps).
• pyrula runs terminate <run_id> ends a run with no worker (Cancel and terminate).
• In a what-if or case run, refund is a write and is refused unless recorded or allowed.

The engine

Build Agents @agent and @tool on the durable engine. Anthropic, OpenAI and LiteLLM clients, MCP, an HTTP surface. Durable Workflows Steps, timers, signals and children, replayed from an append-only journal on SQLite or Valkey. Stream Kafka A pure-Rust client, no librdkafka. Agent output and its offset commit can share one transaction. Core Primitives Rust-backed Either, Option, Try, immutable IList/IMap, serde.

Coming from Pydantic AI, LangGraph, Celery, Temporal or FastAPI? Each page maps that tool’s concepts to Pyrula’s. For coding agents: llms.txt.