Monitor a Claude Code session

Claude Code is the shape Sentisec was built for: many requests per task, each carrying tool calls that touch your filesystem, your shell, and the network. Putting Sentisec in front of it is one environment variable.

You need: Claude Code installed, a Sentisec workspace and API key, and an Anthropic key stored under Settings → Providers, in Provider keys.

Paying for Claude Pro or Max instead of using an API key? Use subscription mode — one command, no key, and you keep running claude exactly as you do today.

Point Claude Code at Sentisec

export ANTHROPIC_BASE_URL="https://api.sentisec.ch/ws/<your-workspace-id>"
export ANTHROPIC_API_KEY="$SENTISEC_API_KEY"
claude

That is the whole integration. Claude Code speaks the Anthropic Messages API with tool_use blocks, which Sentisec already understands — nothing about how you use the agent changes.

Decide what happens on a block

Two settings in the dashboard shape what a monitored session does. Both live under your workspace's Settings.

Protection mode decides whether Sentisec acts or only watches:

• Shadow — every call is scored and recorded, nothing is changed. Start here to see what your sessions actually do before you enforce.
• Enforce — a stopped tool call is replaced before it reaches the agent.

Enforce is the safe default: if the setting is missing or unreadable for any reason, Sentisec resolves to enforce rather than leaving the workspace unprotected.

Egress allowlist names the hostnames and IP ranges this workspace's agents may reach. Entries are hostnames — optionally with a leading * — or IP CIDR blocks. A full URL or a scheme is rejected, so example.com and *.example.com are valid and https://example.com is not.

What a session looks like

A single task spans many decisions:

• Read a file → NOMINAL
• Run the test suite → NOMINAL
• Edit a file → NOMINAL
• Fetch a page the task never called for → ELEVATED
• Pipe a downloaded script straight into a shell → HIJACK

On that last one, in enforce mode, Claude Code receives no tool call: Sentisec withholds it, and an explanation of why takes its place in the answer. The command never executes.

Where to look

Your workspace's Sessions list shows one entry per Claude Code invocation, not one per request — Claude Code sends a session id on every request and Sentisec stitches them into a single timeline. Open a session to see the verdict per step and the tool call that earned it.

Blocked actions also produce signed forensic evidence, retrievable from the session's forensic view.

Next