Tool discovery
An agent sees only the MCP operations granted to its identity and delegation.
Developers · MCP gateway
Connect Claude Code or Codex, and decide each call before a credential is used.
Your own agents, such as Claude Code or Codex, connect to Deixic over MCP (Streamable HTTP). They read the compliance records for your connected systems, and any change they make through Deixic is checked against your written policy before an upstream credential is resolved, then held for a named approver or run and recorded.
Request path
The upstream tool does not run until policy decides.
- 01The agent connects with signed identity and delegation context.
- 02Deixic filters tool discovery to the operations allowed by its grants.
- 03A tool call is evaluated against policy before credential resolution.
- 04The call is allowed, denied, or held for a named approver.
- 05An approved call resumes against the upstream MCP server.
- 06The response is masked where policy requires and the decision and result are recorded.
Controls applied
Identity, capability, policy, and output stay on the same path.
Call decision
Policy can allow the call, deny it, or require approval before an upstream credential is used.
Held execution
A matching approval resumes the parked call; a denial leaves the upstream tool untouched.
Response handling
Recursive masking can remove protected response fields before the result returns to the agent.
Audit evidence
Each call is an action record an auditor can sample.
Durable activity ties the tenant-bound decision to the tool call and its result: which agent asked, what policy decided, who approved with what reason, and what ran. Development recordings can be written as JSONL.
Follow the action lifecycleRead the evidence recordsGateway vs. action control