Context visibility
See what contributes to the estimate
Combines detected agent terminals, manual sessions, open editor context, configured usage files, and locally available model metadata in one dashboard.
AI developer trust tooling ← Software
Agent Inspector
A VS Code-compatible observability layer for understanding what an AI coding session may see, where usage numbers come from, and which trust controls stand between a model request and your codebase.
Context & trust dashboard
Make the signal and its confidence visible
01
Detect local signalsTerminals, installed extensions, open files, and configured logs
local 02
Classify usage provenanceExact, manually recorded, estimated, or unavailable
labelled 03
Review trust signalsPermissions, data flow, project patterns, and policy state
inspectable 04
Gate assisted editsScope, redact, confirm, diff, verify, and retain rollback
controlled 9Documented agent profiles
3Dashboard areas
24Registered commands
NoneExternal background telemetry
Extension capabilities
Move from “the agent is running” to inspectable context and control
Agent Inspector connects session visibility with privacy, scanning, and safer edit mechanics while preserving the difference between observation and inference.
Usage provenance
Exact and estimated stay separate
Labels provider-log or manually recorded message tokens distinctly from approximate open-document context, with source diagnostics instead of false precision.
Extension review
Inspect permission and data-flow signals
Reviews installed extension manifests and bounded source patterns for disclosed capabilities, outbound behavior, sensitive access, and configurable security findings.
Project scanning
Bring policy closer to the code
Scans selected files or folders for sensitive filenames and security-relevant patterns, with path exclusions and importable custom regex rules.
Safe edits
Plan, diff, apply, verify, recover
Constrains AI-assisted edits to selected paths, previews the proposed change, checks generated output, stores a snapshot, and supports one-click rollback.
Outbound policy
Guard the request boundary
Applies domain allow/deny policy, configurable redaction, privacy confirmation, payload metadata, and an outbound audit trail to explicit cloud-assisted edit flows.
Safe Agent path
Put reversible controls around an AI-assisted edit
01
Scope
Confirm the target is inside the user’s selected path locks.
02
Prepare
Apply policy, redaction, destination checks, and confirmation.
03
Review
Present a diff and verify generated output before acceptance.
04
Recover
Keep the previous content snapshot available for rollback.
Engineering signals
What the implementation demonstrates
Trustworthy telemetry design, policy enforcement, privacy-aware model integration, reversible tooling, and honest degraded states.
Observability without invented access
The extension acknowledges that VS Code cannot expose private counters from other AI providers. It uses only locally visible signals and marks estimates as estimates.
Controls wrap the risky operation
Scope validation, redaction, domain policy, confirmation, diff review, output checks, snapshots, and audit events surround the explicit Safe Agent path.
Local-first state and analysis
Dashboard telemetry, scan summaries, policy state, scope locks, snapshots, and audit entries remain in VS Code extension state. Ollama explanations are optional and user configured.
Workspace trust is a real boundary
The dashboard can open with limited support in an untrusted workspace, while filesystem scans, usage discovery, process probing, and edits require a trusted local workspace.
Selected stack
Editor-native observability with optional local reasoning
VS Code Extension APITypeScriptWebview UILocal filesystem signalsTerminal detectionJSON / JSONL logsPolicy engineSecret redactionOllama (optional)VS Code state
Bring observability and restraint to agentic development
Install Agent Inspector from an editor marketplace, inspect the implementation, or reach out about AI developer-tooling roles.