See what your agents tried to do, what scope they asserted, what data they touched, and where governance concerns surfaced.
No API. No account creation. No network calls. Runs entirely locally on your system.
$ pipx install sentience-governorGet Started
Sentience is a local-first runtime for observing AI agent execution. It captures structured events (intent, scope, context, and memory access) at the moment actions occur, allowing teams to trace how agents interact with systems without requiring API calls, accounts, or external services.
pipx install sentience-governor → attach to your agent (MCP / LangChain / Claude Code) → run your workflow → inspect the trace
On your first run, you'll see
Missing intent declarations → Scope mismatches → Unclassified context exchanges → Memory writes worth reviewing
Use in
The Problem
Model-layer safety
Governs what the model says
Agents don't just respond. They read systems, call tools, write data, and trigger workflows.
Sentience
Governs what the agent does
Not a prompt filter. Not a firewall. The governance layer between the model and the world.
Application logging
Captures what happened after
Teams see what happened. Not what the agent believed it was doing.
How It Works
Sentience captures governance events at the moments that matter most in agent execution.
Agent registration
Who the agent is, what version is running, and what it declares about itself.
Intent declaration
What the agent says it is trying to accomplish before it acts.
Scope assertion
What the agent is about to read, call, or write.
Context snapshot
What data moved through the action, with classifications when available.
Memory write attempt
When the agent tries to persist something beyond the current session.
See the Output
$ sentience-cli render ./trace.ndjson [1] REGISTRATION ✓ Agent claude-langchain-demo-v1 [2] INTENT ✓ Objective: 'Look up customer…' [3] SCOPE ✓ READ crm_get_customer [4] CONTEXT ⚠ classifications=[] POL-003 [5] SCOPE ✓ READ crm_fetch_usage [6] CONTEXT ⚠ classifications=[] POL-003
First-run traces often include warnings. That is expected. Sentience makes the gap between intended behavior and actual execution visible.
Integrate in Minutes
Four integration paths. No agent rewrites.
Select your integration
For Python agents already using MCP in production.
from sentience_governor.wrapper.mcp import wrap_mcp_client client = wrap_mcp_client(client, agent_id="your-agent-id")
Coming soon:
MCP registry · Skills distribution · Cursor · Codex · OpenCode
Trust
No data leaves your system
Sentience is a local Python package. It makes no API calls, requires no account, and connects to no external service by default.
Sync is opt-in only. When enabled, it sends aggregated signals — never raw trace content.
Scope is narrow by design
Sentience reads declared intent, scope, and context metadata at the execution boundary. It does not inspect raw payload content.
Payload classification is only possible when you explicitly provide classification hooks. Nothing is inferred from content.
Auditable by design
Governance events are chained in sequence. A gap, reorder, or missing step is detectable — the trace is structured for audit, not just logging.
Apache 2.0. No vendor lock-in. The runtime is yours to inspect, fork, and run.
Start Local. Scale to Control.
Developer → Operator → Enterprise
Developer
FreeOperator
Early accessOpen runtime is free. Control Plane and enforcement are enterprise.