Sentisec
Runtime security for AI agents. Drop-in for Anthropic, OpenAI, Kimi, Gemini, and Vertex AI. Block hijacked tool calls, ship signed forensic evidence.
Documentation
Runtime security for AI agents
Sentisec inspects supported model traffic, applies your workspace policy to proposed tool calls, and records decisions with auditable evidence.
30-second start
Sign up at app.sentisec.ch/sign-up,
create a workspace, connect your provider key, and mint a Sentisec API
key from the dashboard. Export it as
SENTISEC_API_KEY, then change one URL.
curl -X POST https://api.sentisec.ch/ws/$SENTISEC_WORKSPACE_ID/v1/messages \
-H "Authorization: Bearer $SENTISEC_API_KEY" -H "anthropic-version: 2023-06-01" -H "Content-Type: application/json" \
-d '{"model":"claude-sonnet-4","max_tokens":256,"messages":[{"role":"user","content":"hello"}]}'Your workspace id sits next to your API key in the dashboard. Always use the workspace-scoped form above — it is the one URL shape that carries your workspace's policy on every transport. See your base URL.
Need data residency or an air-gapped install? Run Sentisec in your own VPC — enterprise access.
What you get
Use your SDK with a workspace base URL and Sentisec key. Messages, Chat Completions and supported Responses lanes preserve their API shapes while applying the documented denial and streaming contracts.
Inspect recorded decisions using four verdict levels. Read the applied intervention separately to understand what was stopped.
Closed-session evidence with integrity verification and visible failure reporting. See the evidence guarantee and limits under Security.
Claude Code, Cursor, Codex, LangChain, LangGraph, PydanticAI, LlamaIndex, CrewAI, and the OpenAI and Anthropic SDKs.
At a glance
Choose your path
Quickstart
Make your first monitored call in under a minute.
Concepts
Understand the integrity engine before you wire it.
Examples
Working recipes for Anthropic, OpenAI, Claude Code, and LangChain.
Dashboard
Sign up, create a workspace, see your first verdict.
Verdicts
NOMINAL allowed; the steady state. ELEVATED anomalous; logged. SUSPICIOUS off-policy; gated or blocked. HIJACK high confidence compromise; halted.
See Verdicts and interventions for the full mapping.
Providers
Six model providers route through Sentisec today: Anthropic, OpenAI, Amazon Bedrock, Kimi (Moonshot), Google Gemini, and Google Vertex AI. Connect a credential in your workspace and calls dispatch immediately — no restart, no configuration file.
Azure OpenAI is not available yet. If it is a requirement for your deployment, tell us.
Enterprise
In-VPC deployment, configured enterprise sign-in, role-based access, audit logs and SIEM exports. Hosted identity uses Clerk; directory sync and offboarding have explicit boundaries. Self-managed OIDC requires its own provider and workspace configuration.
🔒 Internals
Evaluating the full mechanism, or running the stack yourself? The internal documentation covers it.