A unified harness substrate for agentic frameworks.

Trace, budget, sandbox, auth, and HITL — for any orchestrator. Entorin slides under your agent loop and supplies the harness layer everyone otherwise reinvents.

Quick start Read the guides GitHub

What you get under your loop

One OTel trace per run

Every node, LLM call, tool call, sandbox exec, checkpoint, and handoff is a span — carrying run_id, principal_id, tokens, and cost.

One budget ledger

A single ledger across heterogeneous orchestrators. BudgetGate.check runs before any wire activity; cap-exceeded runs terminate before the network call.

BYOK / multi-tenant

API keys flow as Capability, never read from os.environ. Per-principal scopes for memory, tools, retrieval, checkpoints.

How Entorin differs from agentic frameworks

Frameworks build DAGs. Entorin builds the harness under them. It plugs in; it doesn't replace.

	Agentic framework	Entorin
Workflow shape	DAG / state graph you author	Whatever you already use — bare loop, SDK, framework
Observability	Per-framework, often partial	One OTel trace per run — every span, every adapter
Budget	Reinvented per framework, if at all	One ledger + gate across heterogeneous orchestrators
Auth / keys	os.environ reads scattered across the SDK	Capability flow at substrate boundary; never reads env
Tool & agent protocols	Bespoke abstractions	Reuses MCP (tools) and ACP (agent-as-tool)
What you write	Subclass nodes, register tools, build a graph	Plain Python; substrate emits without your loop knowing

What Entorin does not do

Scope discipline is the substrate philosophy. If you find yourself wanting one of these from Entorin, that is a sign the wrong tool is on your shortlist.

No DAG / workflow builder

That's LangGraph, CrewAI, et al.
No prompt templating

That's your code.
No vector DB / retrieval store

Retrieval ships as a protocol; bring your own backend.
No deployment infra

No Docker, K8s, queues, or load balancers.
No eval suites

Traces are the regression substrate; you bring the assertions.

Quick start

Install the package, mint a RunContext, drop into a bare loop. The full harness emits because the substrate primitives emit them — your loop doesn't need to know.

Read the quick start Browse all guides

# install
uv add entorin

# run
from adapters.bare_loop import run_qa

answer = await run_qa(
    model=anthropic_model,
    ctx=ctx, bus=bus,
    question="how many roads must a man walk down?",
    tool_clients={"fs": fs_client},
    max_turns=6,
)

Pick your path

Claude Agent SDK

Httpx transport + wrapped MCP sessions + per-run sandbox.

Read →

Codex CLI / SDK

Same shape, OpenAI wire format. Chat Completions + Responses.

Read →

LangGraph

Node-wrapper + HITL driver. Substrate inside; LangGraph outside.

Read →

Bare loop

No SDK at all. 50 lines of Python inherit the full harness.

Read →