Sevenbox, find what an AI agent's wrong decision affected

Sevenbox is causal audit infrastructure for AI agent decisions. A Warrant records why a decision was made: the specific prior claims it depended on, and whether acting on it can be undone. Every call your agent makes becomes one, automatically, the moment you wrap it. When a belief your agent acted on turns out to be wrong, Sevenbox finds exactly what still needs fixing — not by searching logs, by walking the causal chain.

This repo is the open-core layer: the SDK, 14 framework adapters, the local Operations Center UI, and examples. It depends on warrant-core, a separate, compiled engine package — not on PyPI, distributed as a GitHub Release of this repo (see Install and License).

The local Operations Center UI (sevenbox ops) ships with this repo — every run your agent makes, with the causal graph above.

Neither package is on PyPI yet. Both come from this repo.

1. Install the compiled warrant-core engine. Go to this repo's Releases page and download the wheel matching your OS and Python version (3.10-3.13), then:

pip install warrant_core-0.1.1-<your-platform-and-python-tag>.whl

Wheels are built for macOS (arm64), Linux (manylinux x86_64), and Windows (AMD64) — e.g. warrant_core-0.1.1-cp312-cp312-macosx_11_0_arm64.whl for macOS on Python 3.12, or ...-manylinux_2_17_x86_64.manylinux2014_x86_64.whl for Linux, or ...-win_amd64.whl for Windows.

2. Install sevenbox itself, directly from this repo:

pip install git+https://github.com/sevenboxlabs/sevenbox.git

Or, if you've already cloned it:

pip install .

pip resolves sevenbox's warrant-core>=0.1.1 requirement against the wheel you installed in step 1 and won't try to reach PyPI for it.

The first time you import warrant_core, you'll be shown its evaluation terms and asked to press Enter to accept. For CI or non-interactive use, set SEVENBOX_ACCEPT_TERMS=1 instead.

from sevenbox_sdk import SevenboxSession, derive, irreversible, oracle, tool

@oracle def infer_intent(ticket): return {"intent": "refund_request", "order_id": ticket["order_id"]}

@tool def lookup_order(order_id): return {"status": "shipped"}

@derive def eligibility(order): return {"eligible": order["status"] == "shipped"}

@irreversible @derive def refund_action(elig): return {"action": "refund_action(4471)"} if elig["eligible"] else None

with SevenboxSession() as session: ticket = session.axiom({"order_id": "4471"}) action = refund_action(eligibility(lookup_order(infer_intent(ticket)["order_id"]))) for w in session.frontier().live: print(w.content)

See docs/quickstart.md for the replay example and a real-framework (LangGraph) walkthrough.

Validated against 14 real agent frameworks, plus the case of running no framework at all (sevenbox_sdk's own decorators, directly — no adapter needed): LangGraph, OpenAI Agents SDK, PydanticAI, CrewAI, Deep Agents, Microsoft Agent Framework, Mastra, OpenHands, LlamaIndex Agents, Smolagents, Letta, Google ADK, AWS Strands Agents, Haystack Agents. A 10-15 line usage page for each lives in docs/adapters/.

Finding a pending compensation obligation is half the story; session.discharge() resolves one, either because a compensating action was actually taken or because it turned out not to need one. Once discharged, it stops showing up as pending. See examples/discharge_obligation.py and docs/api-reference.md for both cases end to end.

222 tests in this repo, covering the SDK and all 14 adapters, pass against the real installed framework packages — no network, no API keys. The full suite, including warrant_core/warrant_runtime themselves (private, not in this repo), is 409 tests.

sevenbox init # writes a real, runnable starter script python my_agent.py sevenbox ops # opens a local Operations Center over whatever store you point it at

More at sevenbox.dev.

The SDK, adapters, and demo in this repo are licensed under Apache 2.0. The warrant-core engine is distributed separately as a compiled binary, under its own evaluation terms (TERMS.md, included with that package) — not open source, not yet licensed for redistribution.