§INTEGRATION · CLAUDE CODE
Claude Code, safely.
Claude Code respects `ANTHROPIC_BASE_URL`. Point it at us and every proposed action runs through the Sentisec control plane before it executes. No code changes in Claude Code itself.
ONE-LINE SETUP
Two lines. One terminal. Done.
Run the installer, then `sentisec login` to pair the workspace. We don't touch your shell rc files silently — you paste the export yourself.
bash
$ curl -fsSL https://get.sentisec.ch | bash $ export ANTHROPIC_BASE_URL=https://api.sentisec.ch/v1
PATH OF A TOOL CALL
One hop. No framework rewrite.
Diagram is intentionally kept at the category level — the detail of the checks lives on the developer landing page and under NDA.
claude (your laptop) │ ▼ proposed actions sentisec control plane │ cognitive integrity score · workspace policy ▼ Anthropic Messages API │ ▼ tool_result back to claude — or halt + reason
§CATCHES
What this catches
- Bash commands whose arguments reference `~/.ssh/`, `~/.aws/`, `credentials*`, `.env`, or known key-file patterns.
- Bash / WebFetch / connector calls whose destination is outside the workspace boundary.
- Shell commands whose content is derived from retrieved web pages.
- Prompt-injection patterns where the proposed action diverges from the user's original task.
§LIMITS
What this doesn't (yet)
- Connector responses are handled conservatively in the current release. Deeper connector coverage ships in a later phase.
- If you attach skills or memory that themselves contain adversarial instructions, treat that as a separate review surface until dedicated static review lands.