Sunday, November 15, 2026

Aegis for monitoring what your AI coding agents actually do on Windows

Aegis for monitoring what your AI coding agents actually do on Windows

Aegis is an open-source, local-first desktop dashboard that gives authorized defenders visibility into AI coding agents like Claude Code and Cursor, observing processes, files and network activity without blocking by default.

Toolantropos17/Aegis — a local-first Electron desktop dashboard for monitoring and reviewing AI coding-agent activity on Windows
CategoryAI agent observability and endpoint monitoring (Electron, Svelte, JavaScript)
Primary UseMonitoring detected agents such as Claude Code, Codex and Cursor — processes, sensitive-file observations and TCP endpoints — plus pre-run Local security reviews of projects and skills
Safe UseDefensive visibility for authorized professionals: observing agents on your own workstations, reviewing code before running it, and exporting audit evidence during sanctioned security reviews
Telemetry NoteMonitoring data stays entirely local with no usage telemetry or cloud sync; endpoint naming uses DNS, update checks contact GitHub, and optional Anthropic analysis sends metadata only on explicit request with your own API key

Aegis, published at antropos17/Aegis, addresses a problem that is becoming acute in every engineering organization that has adopted AI coding agents: those agents run with your credentials, your filesystem access and your network, but you have almost no visibility into what they actually touched. The project's tagline — "Your agents have access. You deserve visibility" — frames it precisely. Aegis is an Electron-based desktop application, built with Svelte and JavaScript, that places observed agent activity into a single dashboard. It is open source under MIT, Windows-first, and the repository currently carries 152 stars with an active CI workflow and a documented beta release cadence.

The current source version is 0.25.0-beta, published as a Windows beta on 12 November 2026 with a signed installer manifest. The scope is broad: process recognition covers 112 distinct agents via 265 process-name signatures, including parent/child tree reconstruction and detection of IDE hosts and extensions. Recognized agents include Claude Code, Codex and Cursor. Windows 10/11 is the primary platform, while macOS and Linux remain explicitly experimental — an honest limitation the README states up front rather than burying in a footnote.

Architecturally, Aegis is a monitor-first design, and this is the most interesting design decision in the project. Default monitoring observes and records; it does not automatically block commands or contain agents. Kill, suspend and resume are manual actions. Monitoring also runs without any agent plugin or subscription — it watches processes, file changes in configured sensitive directories, and TCP endpoints associated with detected agent PIDs, assigning each endpoint an allowlisted, unknown or flagged verdict. For defenders, this separation of observation from enforcement matters: the tool produces evidence and attribution rather than silently mutating agent behavior.

The behavioral analysis layer applies 73 sensitive-path detection rules across eight categories, combined with rolling 10-session baselines per agent, anomaly scoring and sequence evidence. This is a lightweight behavioral-EDR model applied specifically to AI agents — a niche that classic endpoint products do not yet address well, because agent activity looks like ordinary developer tooling. The README is candid that recognition and attribution can be incomplete, which is the right posture for anyone using its alerts as evidence rather than proof.

File-level observation on Windows includes open-handle and Restart Manager observations, and the 0.25.0-beta x64 packages add an experimental Kernel-File read capture. That route requires explicit opt-in: you select a folder in the Events workspace and start capture with manual UAC consent, and it stops on Stop, monitoring pause, suspend or quit. Notably, the tool labels reads as Potential sensitive read based on path rules while explicitly stating the label does not establish malicious intent or change risk scores — a refreshing refusal to oversell heuristic output.

The Local security workspace extends the tool from runtime monitoring into pre-run review. You select a project, skill or agent profile folder and receive static findings, configuration evidence and a diff against a previously saved snapshot. There is also offline report import, which makes the tool usable in a workflow where a reviewer examines an artifact produced elsewhere. Combined with JSON/CSV/HTML/ZIP exports from the investigation workspace, this positions Aegis as useful for audit and compliance documentation, not just live triage.

Claude Code users get a dedicated layer: measured token subtotals from matching local sessions, estimated costs, and CPU/RAM resource charts alongside local LLM probes. The cost figures come with explicit pricing uncertainty flagged in the release notes, and runtime inventory covers WSL distribution names and Docker or Podman container candidates — with explicit freshness caveats and no guest process inspection. These caveats recur throughout the documentation and give the project credibility; the authors consistently distinguish what is measured from what is inferred.

Action controls are opt-in and narrowly scoped, which is where the policy-controlled part of the tool lives. Routes include an exact-execution policy for a single selected CLI action with terminal review, an opt-in Windows Job lifetime route for cleanup of Job-member descendants, and a Windows AppContainer route for one reviewed offline action. The README is explicit that these routes do not intercept other agent tools, arbitrary MCP traffic, or activity outside the configured route. Validating MCP gateways can be configured for selected routes, and a separate Claude hook profile is supported through the Action control workspace.

Privacy architecture is a selling point worth noting for any deployment discussion. Monitoring data stays local, with no usage telemetry and no cloud sync. Endpoint naming performs DNS resolution, update checks contact GitHub, and the optional Anthropic analysis feature sends activity metadata only on explicit request, using your own API key at that service's rates. The SECURITY.md privacy section documents this, and the release-verification workflow — downloading the installer, manifest.json and manifest.json.sig together and verifying offline — is documented for supply-chain hygiene.

Installation from source requires Node.js 24.16.0 or later in the 24.x line, uses Electron 45, and follows the standard pattern: git clone https://github.com/antropos17/Aegis.git, then npm ci and npm start. On Windows, npm run build:sidecar after npm ci enables the native Job/MCP routes. A browser preview via npm run dev serves at http://127.0.0.1:8770 with simulated data, native exports and provider requests disabled — a good way to evaluate the interface before committing to an install.

The 0.25.0-beta release also adds eight interface languages (English, Brazilian Portuguese, Spanish, French, German, Japanese, Simplified Chinese and Russian — the six new ones machine-translated with English fallback), translated anomaly notifications, and locale-aware timestamps for local-security and imported results. Earlier features retained include safer JSON/CSV export replacement, result comparison in local security, and the explicitly configured Windows owner-console route. The beta-readiness record distinguishes release integrity, source checks and runtime scenarios actually tested, which is more transparency than most beta-stage projects offer.

Where does this fit for a professional? For security teams whose developers run coding agents on corporate Windows endpoints, Aegis functions as a host-based observability layer for a threat class — autonomous, credentialed, error-prone automation — that traditional EDR agents treat poorly. It is a defensive instrument: it documents what an agent did, flags sensitive-path contact with honest uncertainty, and lets a human decide whether to suspend, kill or simply keep the audit trail. It is not a sandbox and not an agent firewall; anyone needing hard enforcement must configure the narrow action routes themselves and accept their limited scope. Within those bounds, it is one of the more thoughtfully documented entries in the emerging AI-agent-security category.

Official project repository for antropos17/Aegis.
Download Tool

Educational analysis for authorized security professionals. Use only in controlled, authorized environments.

Sponsored
Share articleFacebookXLinkedIn

Continue exploring

Browse all articles →

0 comentários:

Post a Comment

Note: Only a member of this blog may post a comment.