Technical brief · seven minute read
Interpose installs as a native hook in the agent harness, or as a gateway beside a sandboxed one, and checks every tool call before it runs, with no model in the decision. It runs live on multiple harnesses; Claude Code is the integration we deploy today. This page is written for the person who has to answer for it in production: what you can prove, what you can control, and what you can audit.
Three controls were in the sandbox: the sandbox's own filesystem and network policy, the agent runtime's native guards, and Interpose. Each attempt was probed without a model, so a refusal counts only for the layer that made it.
| Attempt | Sandbox | Runtime's own guard | Interpose |
|---|---|---|---|
| Read a file outside the workspace | Refuses | No gate | Refuses before the call |
| Read a credential file inside the workspace | Allows, the workspace is writable | Advisory, and its message names the way round | Refuses before the call |
| Reach an unlisted host | Refuses at the proxy, after the command has started | Did not fire | Refuses before the command runs |
| Write into Interpose's own install | Allows, a nested read only path under a writable parent is not enforced | Allows | Refuses, in every mode |
The sandbox holds the perimeter: what the process can reach at all. Interpose is the only control on the agent's own install surface, on the workspace contents it may touch, and on the order of operations: a refusal before execution instead of a failure after it. interpose render-openshell-policy writes the sandbox policy from the launch policy and stamps the policy hash the daemon attests on /health, so the two cannot drift.
Interpose registers as a native hook and intercepts the agent at three points. MCP tool calls pass the same pre-execution checkpoint as any other tool call. In the gateway shape the same three checkpoints arrive over hook routes, admitted on a hook token and a per session credential.
A declared task is captured and the session scope, if declared, is set. Drift from the task is scored for a person; scope is enforced.
Every tool call is checked before it runs. This is where blocks happen, so the side effect never lands.
Outcomes are recorded and compared against the baseline for the role.
Every tool call is checked against your policy and allowed or denied before it runs, not flagged afterward. Out of the box the policy protects credential stores, private keys, cloud configs, and system directories. A network allowlist gates tool level requests.
Interpose learns each agent's normal pattern of work and surfaces deviations in reports, for the novel behavior a fixed list will not catch.
Every decision, allowed or denied, is appended to a per agent, hash chained, Ed25519 signed trail anchored by a signed manifest. Alter or delete an entry and verification fails.
interpose scan lists the files an agent could reach, scored by how sensitive each looks by path pattern. Read only. Tighten policy before it matters.
init writes a starter policy, and Interpose recommends tightening from the actions your developers approve and deny. Every control has an alert setting and a block setting, so the security team decides which risks it watches and which it stops. Policy is set centrally and cannot be edited from a developer machine.
# .interpose.yaml
policy:
allow:
actions: [file_read, file_write, tool_invocation, network_request, command_exec]
targets: # logged if outside, denied in block mode
- "src/**"
- "test/**"
networkHosts: # network_request denied unless listed
- "api.anthropic.com"
- "github.com"
forbid:
targets: # hard deny, never released
- "secrets/**"
- "/etc/**"
enforcement:
mode: block # alert runs every check, signs every finding, enforces nothing but self protection
restrictAfter: 2 # eligible violations, per session, inside a 24 hour window
quarantineAfter: 3 # quarantine after this many
scopeBoundarydefault advisoryegressScreendefault offaskOnDestructivedefault onunknownToolsdefault warninterpose --verify-audit re-walks the whole hash chain and checks every Ed25519 signature, then returns a single VALID or INVALID verdict you can hand to a reviewer.interpose --show-audit prints the signed entries behind that verdict, newest first. Paths and commands are redacted to shape by default. A deny names the file that was reached, not the secret it would expose.interpose release from a separate terminal. Interpose reconstructs why the stop fired and proposes a narrow, expiring exception for a person to approve. Nothing is applied on its own.This page reflects the present shipped state and moves as the product ships. Figures elsewhere on this site are illustrative unless marked otherwise.
Thirty minutes to map where your agents run and agree the fixed fee for the audit. Two weeks in transparent mode, so nothing is blocked while you watch, and the report is yours either way.
Questions on anything above? [email protected] or [email protected]. A founder answers.