waxell

Docs · Quickstart · Evaluation walkthrough · CLI Reference · Coming from LangChain/CrewAI?

Build and run governed AI agents. One install gets you the durable execution engine (spawn/suspend, budget enforcement, governance hooks, lineage), the authoring decorators, the wax CLI, and observability. Python 3.10–3.14.

pip install waxell

Using another agent framework and just want observability + governance on top? That's the standalone observe SDK:

pip install waxell-observe   # LangChain, OpenAI Agents, CrewAI, … — no runtime needed

Evaluate it in thirty minutes

waxell-ai/runtime-evaluation is a complete project: one reviewer agent on five frameworks, a domain server on your laptop that verifies every call's end-user and agent identity, a policy that blocks an action, evaluations, memory across four tiers, and signed evidence at the end. The walkthrough explains each step.

What you get

Quick start

An agent is a waxell.yaml manifest — any framework — or Python:

from waxell import agent, workflow

@agent(name="lead_research", description="Research a lead and draft outreach")
class LeadResearch:
    @workflow("draft_email")
    async def draft(self, ctx, domain: str):
        company = await ctx.tool("fetch_company", domain=domain)
        email = await ctx.llm.generate(
            prompt=f"Write a 3-sentence intro for {company['name']}",
            task="outreach_email",
        )
        return {"email": email, "company": company["name"]}

Then push, fire and watch it from the CLI (full command reference):

wax login                                    # authenticate (or: wax config init --profile <name>)
wax doctor                                   # check the install, profile and API reachability
wax push                                     # finds waxell.yaml in this directory and registers it
wax signals fire <signal> --payload '{}' --stream
wax runs                                     # recent runs
wax traces list                              # distributed traces

The CLI at a glance

What's in the box

• Execution engine — ExecutionContext, WorkflowEnvelope, durable Step, checkpoint/resume, ctx.spawn / ctx.spawn_many, tree-budget enforcement.

• Governance — pluggable GovernanceHook, in-runtime policy evaluation before every tool call, lineage.

• Identity — every domain call carries a signed end-user token and a signed agent identity (AXID, on by default) that your systems verify offline against your workspace's public keys.

• Memory — working, session, episodic, semantic and graph tiers declared per slot. Semantic memory embeds with a bundled open-source model by default (no key needed) or any OpenAI-compatible provider you configure.

• Framework adapters — run governed LangGraph, CrewAI, Pydantic AI, AutoGen, and Claude Agent SDK agents (migration guides). Each adapter's framework dependency is optional:

  pip install "waxell[langgraph]"
  pip install "waxell[all]"     # every adapter
  

• Authoring — agent, workflow, router, tool, policy decorators, spec types, and typed Context objects for IDE autocomplete.

• Observability — every run traced; cost, tokens, and policy decisions in the Waxell dashboard via waxell.observe.

Import paths: from waxell import agent is the documented path; from waxell_runtime import agent works too and is kept forever. The internal waxell_sdk package bundled in this wheel is an implementation detail — never import it directly.

License

waxell is distributed under the terms of the MIT license.