Integrations
EVE + Generic Python (Supported)
Compatibility: SUPPORTED — the generic Python adapter is the validated, in-tree integration surface. It routes a tool call through the same governance composition root used by the rest of EVE: CoreGuard makes a deterministic ALLOW / BLOCK / MODIFY decision before the governed action executes, and every decision can produce a signed evidence record.
Use the generic adapter when you control the code that invokes a tool and can wrap the call site. It is framework-agnostic: any Python function you treat as a "tool" can be governed.
What you get
- Deterministic policy decision (ALLOW / BLOCK / MODIFY) before the tool executes.
- No LLM in the CoreGuard decision path — the verdict is computed by deterministic policy code.
- A signed decision-evidence record you can verify independently, offline, without trusting the EVE server (ECDSA P-384 (AWS KMS) in production configuration; HMAC-SHA256 fallback is symmetric and not independently verifiable).
- Fail-closed behavior: if the governance service errors, the SDK does not silently fall back to permissive behavior.
Install (pilot / private)
The eve-coreguard package (version 0.2.7) is published on public PyPI
(pip install eve-coreguard). Its primary entry point is the
CoreGuardClient class:
from eve_coreguard import CoreGuardClient
The pip client governs via hosted mode and needs a reachable endpoint. In-process embedded governance runs inside the EVE service, not the client wheel.
Govern a tool call
from eve_coreguard import CoreGuardClient
client = CoreGuardClient(api_key="eve_sk_...")
result = client.evaluate(
request_id="req-2f9a",
tenant_id="tenant_demo",
proposed_action={"type": "issue_refund", "account_id": "acct_1042", "amount_cents": 5000},
model_output={"decision": "approve", "confidence": 0.9},
context={"principal_id": "agent_support_7", "session_id": "sess_2f9a"},
policy_set="refund_v1",
include_evidence=True,
)
# evaluate returns an EvaluationResult (decision.status / decision.action / risk / ...).
if result.decision.status == "BLOCKED":
# the underlying tool is not called
print(result.decision.action, result.risk.level)
else:
refund = issue_refund(account_id="acct_1042", amount_cents=5000)
On BLOCK, the underlying tool is not invoked, so a blocked call produces no unauthorized side effect. On MODIFY, forward the modified arguments returned in the decision.
Verify the evidence
from eve_coreguard import verify_decision_record
# The signed evidence record for the decision is carried in result.signed_governance
# (evaluate was called with include_evidence=True). For an ECDSA P-384 record, pass the
# published public key so the signature is checked offline, with no shared secret.
report = verify_decision_record(result.signed_governance, public_key_pem=public_key_pem)
assert report.valid is True
Verification proves the evidence is authentic and unaltered and that it matches its schema. It does not attest that the underlying decision was correct — only that the record is genuine.
Readiness
- Generic Python adapter: SUPPORTED — validated in-tree (ALLOW executes, BLOCK causes zero side effects, direct-import bypass guarded).
- EVE SDK core: RELEASE CANDIDATE WITH EXCLUSIONS.
- Python SDK
eve-coreguard0.2.7: release-candidate core client, PILOT_READY. - CoreGuard deterministic evaluation and signed evidence: PILOT_READY / SUPPORTED per the claims registry.
Limitations
- The wrapper governs the call site you wrap. A caller who invokes the underlying tool function directly, bypassing the wrapper, is not intercepted. For a boundary that a direct call cannot route around, deploy a sidecar or restrict the tool registry so tools are only reachable server-side through the governed path.
- The pip client governs via hosted mode and requires a reachable endpoint; embedded in-process evaluation is the EVE service, not the client wheel.
- Independent (asymmetric) verification requires the ECDSA P-384 public key. An HMAC fallback is symmetric and not independently verifiable.
- Signed evidence proves authenticity and integrity of the record, not the correctness of the decision itself.
- This page describes governance of tool calls; it is not a compliance certification.