The gateagent action→deterministic gate→signed record

Every agent action runs one gate your security team controls.

Interpose puts one non-bypassable control between your Claude Code agents and everything they can touch. Deterministic at the core, no model in the block path, running inside your own boundary.

01 · The decision path

Checked before it runs. No model in the block path.

Every proposed action passes a deterministic pre-execution gate before it runs. In transparent mode the gate records and takes no action. In block mode out-of-bounds actions stop before the side effect lands.

agent action
proposed, not yet run
→
deterministic gate
block or allow
→
escalate
gray area only, your model
→
signed record
tamper-evident
Layer 1
Role boundary
Each agent is scoped to a role. Actions outside it are flagged in transparent mode, denied in block mode.
Layer 2
Intent alignment
Each action is measured against the stated task. Intent drift surfaces the moment behavior stops matching it.
Layer 3
Behavioral baseline
Deviation from the role's known pattern surfaces drift a static rule misses. The baseline flags, it does not block.
Layer 4
Pre-execution gate
The deterministic hard stop. Blocked means it never ran. Credentials are denied here by default.
Escalation stays in your boundary
Ambiguous cases can escalate to a model you bring: Anthropic, Gemini, or an open source endpoint. It runs on-prem and nothing reaches Tuent. Credentials never escalate. They are blocked deterministically.
Zero runtime cost on the core path
The gate runs in the same process as the agent and finishes in well under a millisecond. No network hop, no tokens on the block path.
One action, through all four layers
agent reads .env.production → role boundary: outside scope → intent: drift from the stated task → baseline: never seen in this role → gate: denied before it ran · signed record written
02 · Policy control

The security team owns policy. Interpose drafts it.

Interpose watches what your agents actually do and drafts a specific rule in plain language. The security team reviews and approves before anything is enforced. Developer approvals and denials are signal, not authority. Policy is central and the endpoint cannot override it. Every change is signed to the record.

interpose · policy review · payments-api
Drafted from 28 approvals and 3 denials this week. Security approves to enforce.
allowread src and docs26x
allowedit src19x
denyread .env and secretshard rule
escalateunlisted hosts to your model
Approve policy
Adjust
03 · Coverage

What the gate sees and controls.

Every action in the agent's path, evaluated in process at the moment it happens, before the side effect lands.

+Tool and file calls. Every read, write, and tool invocation the agent makes.
+Shell commands. With a guard on irreversible, destructive ones.
+Network egress. Checked against an allowed-host list before data leaves.
+MCP calls and skills. Local and hosted, evaluated like any other action.
04 · Integrations

Live on Claude Code. More runtimes on the roadmap.

Interpose ships today as a native Claude Code hook. Central control installs once and every action routes through the gate. It is the only live integration today.

claude-codelive
next adapterin design

Put every agent action under one control.

Book a walkthrough and we will run a live agent through Interpose: transparent mode, block mode, and the signed record, on your use case. Free three week pilot, shaped to your team, then priced to your fleet.

Book a demo How the pilot works