raxit 0.0.4

Zero-dependency Python client for the raxIT runtime governance engine (permit / defer / deny on every tool call)

raxit — thin Python client for the raxIT governance daemon

Govern every agent tool call with permit / defer / deny, before it executes.

pip install raxit     # zero runtime dependencies (stdlib only); or: uv add raxit

Docs: docs.raxit.ai.

Platforms: the launcher resolves/downloads engine binaries for darwin/linux only. On Windows, run raxit from WSL — there is no native Windows engine build.

Zero-code governance (raxit run)

The fastest path is no SDK calls at all. raxit run boots the daemon on localhost:8181, sets the child environment, and preloads the auto-patcher (via sitecustomize) so every tool call is governed. raxit init scans your repo and generates the .raxit/ policy workspace first — raxit run fail-closes without it:

raxit init                     # scan -> generate .raxit/
raxit run -- python agent.py   # inherits the agent id from .raxit/security.yaml

Under the hood this sets RAXIT_AUTOLOAD=1 and puts the raxit sitecustomize dir on PYTHONPATH, so the auto-patcher fires at Python startup and patches the tool-dispatch choke point of each detected framework:

Framework	How it's governed
LangChain	Auto-patched (importable)
LangGraph	Auto-patched
CrewAI	Auto-patched
AutoGen	Auto-patched
OpenAI Agents SDK	Auto-patched
Pydantic AI	Auto-patched
Google ADK	Auto-patched
LlamaIndex	Auto-patched
Strands Agents	Auto-patched
smolagents	Auto-patched
Haystack	Auto-patched
Agno	Auto-patched
Claude Agent SDK	Auto-patched (injects a PreToolUse hook)
Microsoft Agent Framework	Auto-patched (injects a FunctionMiddleware)
DSPy	Auto-patched (Tool.__call__/.acall for dspy.ReAct, plus ToolCalls.ToolCall.execute for native tool calling — two independent surfaces)
CAMEL-AI	Auto-patched (FunctionTool.__call__/.async_call — the chokepoint every ChatAgent tool call, sync or async, normalizes through)
Agency Swarm	Auto-patched (inherited — tool dispatch, incl. its own SendMessage, flows through OpenAI Agents SDK's already-patched chokepoint; no dedicated patcher)

A blocked call raises RaxitDenied; a MODIFY verdict runs the tool on the engine's transformed arguments; a permit runs it unchanged. raxit.agent also exposes governed / governed_tool as a manual, per-tool escape hatch when auto-patch doesn't fit.

Programmatic client

from raxit import RaxitClient

client = RaxitClient("http://localhost:8181", agent_id="my-agent")
d = client.govern(resource="db.read", args={"query": "select 1"})
if d.deferred:
    d = client.govern.approve_wait(d)      # blocks until a human approves
if d.denied:
    raise RuntimeError(f"{d.reason_code}: {d.hint}")
result = run_tool()                        # only runs on permit
client.output.record(result, car=d.car)    # feeds output governance + taint

LangChain (no langchain import required by this package):

from raxit.langchain import RaxitCallbackHandler
handler = RaxitCallbackHandler(client)    # raises RaxitDenied on deny
agent.run("...", callbacks=[handler])

Fail-closed by default: if the daemon is unreachable, govern() returns deny (deny.daemon_unreachable). Pass fail_open=True for dev only.

RAXIT_GOVERNANCE

Governed by default. RAXIT_GOVERNANCE=off is the kill-switch for contrast/before-after demos — the client then bypasses the daemon entirely and every call synthesizes permit. Only two values are accepted (after strip().lower()): on and off; unset defaults to on. Any other value — 1/true/yes/0/false/no, or a typo — raises ValueError at client construction, before any network call, rather than silently guessing.

Audit — verify the tamper-evident decision log

client.audit.verify() checks the hash-chained Decision Provenance log; pass signatures=True with one or more pinned Ed25519 public keys to also verify record signatures offline against a trusted key (not the daemon's own config):

client = RaxitClient("http://localhost:8181", operator_token="<operator token>")
report = client.audit.verify(
    signatures=True,
    require_signed=True,
    pubkeys=["<base64 raw-32-byte Ed25519 pubkey>"],
)
assert report["signatures_verified"] and report["signed_count"] > 0

Operator-gated (Authorization: Bearer <operator_token>). For fully offline, adversarial-grade verification use the CLI: raxit audit verify --signatures --pubkey <key>.

Budget — read the session/principal budget projection

client.budget.get() is a strictly read-only projection of the engine's budget accountant — never a reservation. It reports used/remaining against each configured cap (call budget, per-session and per-day token/cost caps):

client = RaxitClient("http://localhost:8181", operator_token="<operator token>")
snap = client.budget.get(agent_id="payments-bot", session_id="sess-1")
for b in snap.budgets:
    print(b.name, b.used, "/", b.limit, b.unit, f"({b.window})")

Operator-gated (Authorization: Bearer <operator_token>). Fails closed: an unreachable daemon raises GovernanceError — it is never read as an empty or infinite budget. An empty snap.budgets means genuinely no cap configured.

Warrants — inspect the loaded standing grants

client.warrants.list() / client.warrants.get(id) are a read-only view of the warrants loaded at raxit serve startup. There is no create/revoke — minting stays the operator-controlled load path.

client = RaxitClient("http://localhost:8181", operator_token="<operator token>")
for w in client.warrants.list():
    print(w.id, w.agent_id, w.tool_pattern, w.status, "expires", w.expires_at)

Operator-gated. Fails closed (raises GovernanceError when unreachable, so "none" is never confused with "unreachable"). Warrants carry no credentials and the attested intent-baseline text is never projected.