Your GPU isn’t an all-you-can-eat buffet. A website should not get free inference from your machine.
Secure Agent helps you control what coding agents can read and send, and who can use your local models. Configure file-access guards, outbound-action checks and secret inspection for supported workflows, then review their activity in one place.
- Download
SecureAgent-<version>.dmgfrom the latest release. - Drag Secure Agent.app to Applications and launch it. Requires macOS 14+.
- Choose your harness in Setup. Register Claude Code, Cursor, Codex, OpenClaw, Antigravity, Hermes or OpenCode v2 hooks, or continue with the selected harness's observation path.
- Start an agent session and check its observed coverage. Enable file telemetry, guard prompts, traffic routing and local analysis separately as needed.
Follow Getting started, then Your first observed session to see evidence and export a report. For problems, use Troubleshooting. To build from source, use Development.
These docs describe the current main checkout. The published release may have an earlier setup flow or fewer features; consult its release notes.
- Protect credential files. Set supported file-guard rules to ask or deny when an agent reaches protected paths. File protection
- Control outbound actions. Review recognized sends after untrusted content. Inspect routed requests for registered secrets and known credential patterns, and choose which matching requests to block. File and egress protection · Firewall limits
- Keep control of local inference. Use private MLX analysis or review hardening for supported MLX-LM, llama.cpp and Ollama services. Local model protection
Named guard and firewall rules start in monitoring mode. Choose prompt or deny modes for guard checks, and blocking for selected firewall rules.
Secure Agent also offers session monitoring, resource controls, cleanup and fleet views. Its native macOS menu bar app, local web console, Go daemon and harness hooks work together to show activity and decisions.
Screenshots use synthetic demonstration data.
| Goal | Guide |
|---|---|
| Review findings and incidents, tune alerts, inspect session history | Console and CLI |
| Gate sensitive file access, register secrets, inspect routed payloads | File and egress protection |
| Inspect session memory/CPU and configure resource budgets | Resources |
| Find worktrees, review cleanup and track reclaimed disk space | Worktrees and cleanup |
| Run private MLX analysis or harden existing local model servers | Local model protection |
| Chat with local Ollama and confirm proposed commands | Local chat and commands |
| Collect signed events and posture from several nodes | Fleet |
Open Monitoring in the console to review each running session’s file guard, recorded activity, and traffic inspection evidence. Select a session to inspect its observations, or expand Other detections for processes without a linked session. Review issues explains monitoring gaps; unsupported capabilities and quiet sessions are shown separately from failures.
File guarding, system observation and payload inspection are separate capabilities. Claude Code, Cursor, Codex, OpenClaw, Antigravity, Hermes and OpenCode v2 have supported guard hooks; other harnesses have varying process and transcript coverage. See the harness coverage table.
Named guard rules and the firewall default to monitoring. Some hook safety checks always deny credential-printing and guard/harness mutations. Guard prompts, firewall blocking and automatic resource interventions require separate configuration. Unrouted or tunneled traffic is not payload-inspected, and a manual hook check does not prove a live session invokes it. Read the protection guide and limits before relying on enforcement.
macOS is the primary platform for the app, Endpoint Security telemetry and DMG distribution. The daemon, CLI and collector also build on Linux with CGO_ENABLED=0, using /proc for process/network sampling. See Linux and headless setup.
The documentation index links every guide and reference, including Configuration, API, Architecture and the threat models.
For development and pull requests, read Contributing. Report vulnerabilities through the security policy.
Distributed under the MIT License.













