Your agents have access. You deserve visibility.
AEGIS puts local AI-agent activity in one desktop dashboard. Follow detected Claude Code, Codex, Cursor and other agent processes, investigate sensitive-file observations and network connections, and review a project or skill before you run it. Open source. Windows first. No subscription for monitoring.
Download the Windows beta → · Explore AEGIS · See the interface
Interface preview with simulated data, captured 11 October 2026. Screenshots and provenance.
- Follow the agent behind the activity. Explore detected processes, parent/child trees, file observations and TCP endpoints in one place.
- Investigate sensitive activity. Open an alert, examine its evidence and attribution, then keep a report for later review.
- Check a project before the next run. Review a selected project, skill or agent profile for static findings, configuration evidence and changes since a saved snapshot.
- Understand Claude Code usage. Inspect measured token subtotals from matching local sessions and estimated costs alongside resource charts.
- Add controls where you choose. Configure explicit action policies and validating MCP gateways for selected routes. Scope and setup.
Monitoring runs without an agent plugin. It observes and records; automatic blocking requires a separately configured supported route. Recognition and attribution can be incomplete. Windows is the primary platform; macOS and Linux remain experimental. Capabilities and limits.
Install · Use AEGIS · FAQ · Roadmap · Documentation · Contribute
Current source version: 0.24.0-beta Each installer contains the features at its release tag. Verify your download · Beta readiness
For a first desktop session: install a verified beta, open Monitoring while your coding agent runs, then select an agent to inspect its observed activity. For a review before running unfamiliar code, open Local security and select the project folder. No provider API key is needed for these workflows.
| Your next step | Where to go |
|---|---|
| Explore the interface | View screenshots without installing, or run the browser preview from source. Both use simulated data. |
| Use the Windows desktop app | Choose a published release and verify its installer, then follow the desktop workflows. |
| Help improve AEGIS | Find a contribution by skill, check active PRs for overlap, or ask a question. |
| Area | Coverage |
|---|---|
| Processes | 112 agents (265 process-name signatures), process trees, IDE-host and extension detection; Windows primary, macOS/Linux experimental |
| Files | Changes in configured sensitive directories and agent config paths; Windows open-handle and Restart Manager observations |
| Network | TCP endpoints for detected agent PIDs, endpoint names and allowlisted / unknown / flagged verdicts |
| Behavior | 73 sensitive-path detection rules across 8 categories, rolling 10-session baselines, anomaly scoring and sequence evidence |
| Resources and usage | CPU/RAM charts, local LLM probes, Claude Code token subtotals and estimated costs |
| Runtime inventory | Running WSL distribution names and local Docker/Linux-native Podman candidates; explicit freshness, no guest process inspection |
| Local review | Project, skill and profile inventory; static findings, saved snapshot comparison and offline report import |
| Investigation | Agent instances, file/network views, alerts, audit history, statistics and JSON/CSV/HTML/ZIP exports |
Monitoring requires no agent plugin or subscription. Default monitoring observes and logs; it does not automatically block commands or contain agents. Kill, suspend and resume are manual actions. See the full capability guide.
Download the installer, manifest.json and manifest.json.sig from the same
GitHub release, then follow
offline verification. Release notes describe that
installer's features.
Install the first beta manually after verification: the published 0.17.0-alpha updater cannot discover beta releases. When upgrading from 0.14.1-alpha or older, install 0.15.0-alpha or newer manually before using signed in-app updates.
Requires Node.js 24.16.0 or later in the 24.x line. Source uses Electron 45 and 64-bit platforms; macOS requires version 13 or later. Windows 10/11 is primary; macOS/Linux remain experimental.
git clone https://github.com/antropos17/Aegis.git
cd Aegis
npm ci
npm startOn Windows, run npm run build:sidecar after npm ci for native Job/MCP routes
and native-helper tests. Linux process actions additionally require Python 3.9+
at /usr/bin/python3 and kernel pidfd support; see Linux process control.
The development workflow covers helper builds and checks.
From a source checkout after npm ci, run:
npm run dev
# open http://127.0.0.1:8770The preview shares desktop components and uses simulated data with an isolated
host. It disables native exports and provider requests. npm run frontend:build:preview
builds that preview; npm run build:renderer builds the desktop renderer.
Source opens Advanced navigation by default. Simple interface in Settings offers investigation, file checks and settings; a saved Simple choice is retained.
| Task | Start here |
|---|---|
| See agents and their activity | Monitoring, then an agent's processes, files or connections |
| Review a sensitive-file alert | Alerts, then inspect the captured evidence |
| Inspect a Windows file read | Events → Folder to monitor → Start file read monitoring; setup and limits |
| Review a project, skill or profile | Local security |
| Check a selected action against a policy | Action control |
| Review a separate Claude Code hook profile | Action control → Separate Claude hook profile; setup, supported launch and removal |
| Configure selected-action MCP tools | Terminal setup |
| Find other tools | Start here or Commands (Ctrl K); workspace guide |
| Monitoring | Local security |
|---|---|
 |
 |
These are simulated source UI captures from 11 October 2026. All screenshots and provenance.
Windows x64 packages include experimental Kernel-File reads. In Events, choose a local folder and explicitly start capture with manual UAC consent. It stays off on startup and stops on Stop, monitoring pause, suspend or quit. Events can display a read, its verified issuing agent and Potential sensitive read from path rules. Paths remain candidates, and the label does not establish malicious intent or change risk scores.
Monitoring data stays local, with no usage telemetry or cloud sync. Endpoint naming uses DNS; update checks contact GitHub. Optional Anthropic analysis sends activity metadata on request using your own API key and incurs that service's charges. See privacy and key handling.
These routes require explicit setup and cover selected actions only. They do not intercept other agent tools, arbitrary MCP traffic or activity outside the route.
| Route | Scope |
|---|---|
| Exact execution policy and terminal review | One selected CLI action; policy deny and failed preparation do not launch. Fresh confirmation can be required per attempt |
| Windows Job lifetime | Confirmed cleanup of ordinary Job-member descendants; caller file/network privileges are retained |
| Windows AppContainer | One reviewed offline action in a new retained workspace, zero capabilities, private-file access checks and Job cleanup |
| Selected-action MCP tools | Up to eight fixed argument-free tools, optional review broker, configuration checks and connection status |
| Validating MCP gateways | Explicit stdio, loopback HTTP and pinned HTTPS; accepted catalogs and bounded grants |
| Reviewed file deletion | One exact AEGIS-owned file operation with mandatory terminal confirmation |
Optional persistent grants, known-secret checks and live route observation have separate contracts. Configuration checks grant no execution permission; observation does not verify provider identity or blocking. Client settings stay under operator control. Terminal previews and generated configurations can contain private paths or secrets, so keep them private. Setup guides.
Administrative repository APIs can create a private Hyper-V VM, hydrate its original disk from a trusted Windows VHDX, and run one bounded executable as a fresh standard guest user. Actual boot, quota refusal, descendant cleanup and VM removal passed on the recorded source.
The guest has no NICs, a private desktop, a 512 MiB allocation quota, a
32-process/512 MiB Job and a 120-second run limit. These require an administrative
disposable lab. Ordinary installed-caller launch stays disabled
(launchAllowed=false); complete session integration and independent A1–A4
acceptance remain open. Administrative contract
· Protection roadmap.
The immediate target is a usable Windows Claude Guard: install, launch the supported route, allow ordinary work, block declared dangerous attempts and inspect the result. This is an active delivery target; full installed-route acceptance remains open. See the execution order and current evidence and gaps.
The completion plan tracks the current work and its acceptance boundaries. Source changes and the development context identify the active checkout and evidence. Each qualification record names its own tested revision and package; a source change does not update a published installer.
Recent Observatory work improves keyboard navigation, focus, localization and the clarity of observation and settings states in #1127 and #1128. These source changes do not establish installed-application or native-sensor acceptance.
The following is a historical snapshot from 18 October 2026, through
#975 at
96b63b16.
Read the engineering milestones and dated native qualification evidence
| Area | Merged work |
|---|---|
| Development context | #960 aligns context and skills; #962 adds Windows fixture CI. Contributor workflow |
| Packaged monitoring | #963 checks the built window and Start/Query/Stop; #964 fixes startup freshness. Evidence boundary |
| File and TCP sensors | #965 refreshes TCP observations after settings changes and checks owned file/socket activity. Packaged journey |
| Diagnostic ETW | #966–#973 test scoped reads and recovery after helper, main and custodian loss. Recovery limits |
| Genuine agent read | #975 verifies installed Claude Code reading a synthetic .env after a preparatory README read, with a visible read, issuer and potential risk. Live record |
#974 was closed without merging. Its preflight did not establish an actual Windows restart and post-restart capture; see the current Windows recovery admission.
#976 prepares detected threads before the first scoped read. Its live qualification passed the genuine first-read journey without README warmup and the persistent-read regression on one diagnostic Windows package. Both showed a verified agent read and potential risk; additional observations can remain unverified. This candidate was tested with the diagnostic profile. Ordinary packaged controls have their own setup and qualification.
The idle first-read qualification passes genuine Claude Code reads after 65, 70 and 85 seconds in an ordinary package, with verified issuer, visible action/potential risk and cleanup. It covers renewed original handles; the earlier attribution failure is retained. Collector drops remain separate from kernel loss; these passes do not establish complete capture or attribution of every worker.
- Coverage: Unknown signatures and short-lived processes can be missed. Default monitoring does not parse individual tool calls. Healthy sensors or a valid audit chain do not prove complete capture.
- Platform identity: macOS lacks a process-birth witness. Linux depends on accessible /proc data; fallback lacks a start-time witness. Missing identity limits attribution and process actions.
- File evidence: Candidate paths do not resolve reparse targets. A first read, new worker or unavailable native witness can remain unverified. Diagnostic reads do not enter risk scoring, FileEvents, audit, baselines or sequences.
- History and recovery: UI windows are bounded; Statistics exposes eviction counts and Audit retains persisted records. File-watch retries and ETW recovery have bounded, scenario-specific limits; observation gaps remain explicit.
- Token usage: Claude Code requires a matching process lifetime and readable session data. Bounded reads, oversized records and index capacity can leave subtotals incomplete; other-agent values and dollar amounts can be estimates. Accounting and soak limits
- Privacy: Paths, agent names and endpoints can appear in logs and exports. API-key storage depends on a secure safeStorage backend; security guidance explains the boundary.
- Unmeasured behavior: No overall detection rate, false-positive rate, startup-time guarantee or whole-app overhead figure has been established. Static findings and scores do not establish safety.
Help complete the Guard journey through setup documentation, reproducible failures, meaningful regression cases and decision feedback. The contribution guide gives bounded starting points and their acceptance conditions.
Choose an open good first issue for a small, scoped change, or browse help wanted for work that needs more context. The task picker connects documentation, Svelte/TypeScript UI and JavaScript/Vitest regression work to starting files and focused issue queues. Read the issue's expected result and platform needs before starting; small contributions do not require provider credentials.
Start with CONTRIBUTING.md, the current development context and code conventions. Main uses CommonJS JavaScript; the Svelte renderer and shared types use TypeScript. Exact versions and scripts are in package.json.
Use an isolated branch, preserve existing work and describe actual validation in
your PR. npm test runs the configured suite; npm run counts:check checks
maintained declarations. CI also runs builds, lint, type checks, dependency audit
and derived counts. Mutation campaigns are not part of routine CI. Native and
provider scenarios need their own matching evidence. Architecture
· Roadmap.
Help more developers see what their agents are doing. If AEGIS is useful to you, star the repository, share a workflow that helped, or tell us what made setup difficult in Discussions. Reproducible feedback helps shape the next beta.
For setup questions, start with the FAQ and user guides, then ask in Discussions. For a reproducible bug, use the bug-report form and include your AEGIS version, whether you use an installer/source/preview, your OS and the affected workspace. Remove private paths, file contents and credentials before sharing logs or screenshots.
Bug report · Feature request · New agent submission · Discussions · Private vulnerability report
MIT license · Code of conduct · All contributors. Navigation icons use a curated Tabler subset with its MIT notice and pinned provenance.