EVE + Claude Agent SDK
Compatibility: Experimental integration — structurally implemented but not included in the supported release-candidate adapter matrix.
The Claude Agent SDK adapter routes an agent's tool calls through EVE's governance composition root so CoreGuard can make a deterministic ALLOW / BLOCK / MODIFY decision before a tool executes. The adapter is structurally implemented, but the Claude Agent SDK was not installed during release-candidate validation, so it was not validated against a pinned live version and is marked UNTESTED / experimental. Only the generic adapter is SUPPORTED.
How governance works here
Each requested tool call is submitted to EVE before execution. No LLM is in the decision path.
from eve_coreguard import EVE
eve = EVE() # hosted endpoint
decision = eve.govern_tool_call(
tool="write_file",
arguments={"path": "/workspace/report.txt", "contents": "..."},
context={
"tenant_id": "tenant_demo",
"principal_id": "agent_writer",
"session_id": "sess_a17c",
},
)
# ALLOW -> execute; BLOCK -> do not call the tool; MODIFY -> use modified arguments.
eve-coreguard (0.2.2) is distributed privately during a pilot and is not on a public index.
Entry point: from eve_coreguard import EVE.
Hardened alternative
This adapter is wrapper-enforced: a code path that calls the underlying tool directly is not intercepted. For a boundary a direct call cannot route around:
- Put tools behind a restricted server-side tool registry so the agent reaches them only through the governed path.
- Proxy tool execution through a sidecar so governance runs out-of-process.
- Apply network enforcement so tool endpoints are only reachable via the governed path.
- For MCP tools, route through the MCP gateway, which binds the approved request to the executed request (see mcp.md).
Readiness
- This adapter: experimental (UNTESTED against a pinned live Claude Agent SDK version; wrapper-enforced).
- Generic Python adapter: SUPPORTED.
- EVE SDK core: RELEASE CANDIDATE WITH EXCLUSIONS.
Limitations
- Not validated against a pinned live framework version.
- Wrapper-enforced: bypassable by a direct underlying-tool call. Hard enforcement requires a gateway/sidecar or restricted server-side tool registry.
- If you are integrating via Claude Code hooks specifically, note that those hooks are cooperative and are not an enforcement boundary — see claude-code.md.
- Hosted mode requires a reachable endpoint; embedded governance is the EVE service.
- Evidence verification proves authenticity and integrity, not decision correctness; independent verification needs the Ed25519 public key.