interven-computer-use

Govern Anthropic Computer Use tool calls through Interven before they run on your host.

Computer Use tools — computer_* (mouse/keyboard/screenshot), bash_* (shell), text_editor_* (file edits) — execute locally and never touch HTTP. The only place to govern them is the agent's own loop: after messages.create returns tool_use blocks and before you execute them. IntervenGuard sits exactly there. Each tool call is scanned through Interven's policy + risk engine, and the decision is mapped back into Computer Use's allow/deny model.

Install (30 seconds)

pip install interven-computer-use
export INTERVEN_API_KEY=iv_live_...   # from https://app.intervensecurity.com/signup

Use

Wrap the tool-execution step of your loop. The guard scans, then either runs your executor (ALLOW/SANITIZE) or returns a synthetic error block (DENY/REQUIRE_APPROVAL):

import anthropic
from interven_computer_use import IntervenGuard

client = anthropic.Anthropic()
guard = IntervenGuard()  # reads INTERVEN_API_KEY from env

def run_tool(block):
    # your existing executor: screenshot, run bash, edit a file, ...
    return execute_computer_use_tool(block.name, block.input)

messages = [{"role": "user", "content": "Open the terminal and run `whoami`."}]

while True:
    response = client.beta.messages.create(
        model="claude-sonnet-4-5",
        max_tokens=1024,
        tools=[
            {"type": "computer_20250124", "name": "computer",
             "display_width_px": 1280, "display_height_px": 800},
            {"type": "bash_20250124", "name": "bash"},
            {"type": "text_editor_20250124", "name": "str_replace_editor"},
        ],
        betas=["computer-use-2025-01-24"],
        messages=messages,
    )

    tool_results = []
    for block in response.content:
        if block.type == "tool_use":
            # One line of governance — scan + execute-or-block:
            tool_results.append(guard.guard_tool_use(block, execute=run_tool))

    if not tool_results:
        break  # no tools requested — Claude is done

    messages.append({"role": "assistant", "content": response.content})
    messages.append({"role": "user", "content": tool_results})

Prefer to make the decision yourself? Call guard.check(block) and inspect the Verdict (.proceed, .decision, .reason, .approval_url, .trace_id).

Decision mapping

A blocked call returns a standard Computer Use tool_result block:

{ "type": "tool_result", "tool_use_id": "tu_…", "content": "Interven DENY (…). Trace: …", "is_error": true }

Configuration

All of these can also be passed to the constructor:

guard = IntervenGuard(
    api_key="iv_live_...",
    gateway="https://interven.your-company.internal",
    timeout=10.0,
    fail_closed=True,
    agent_name="ops-desktop-agent",
)

Failure mode

If /v1/scan is unreachable (network down, gateway 5xx, timeout, missing key), the guard fails open by default: the tool runs and a warning is written to stderr, so a partial outage doesn't freeze the agent. Set INTERVEN_FAIL_CLOSED=1 (or fail_closed=True) to block instead — choose this when enforcement matters more than availability.

What gets scanned

Every tool_use block you pass to guard_tool_use / check is scanned — not just the three built-in families — so a custom tool your agent calls is governed too. The guard infers a category (computer / bash / text_editor / other) from the tool name and includes it in the scan so policies can match on it. No tool args are executed before the decision returns.

License

MIT