Built for Governed Enterprises
AgentMux is the open-source AI agent control plane built around local-first governance. Agents run on your machine, under per-agent identities, with no telemetry and nothing phoning home. The code is Apache 2.0 and auditable.
Security Architecture
AgentMux sits at the interaction layer, the point your AI agents route through. Only agents you register can run, so isolation and accountability come from the architecture rather than a layer added after the fact.
Zero Telemetry
No analytics, no phone-home, no data sent to AgentMux servers. The source is open, so you can confirm it.
Local-First Data Path
Agent activity stays in your user profile by default, air-gap capable. An opt-in cloud relay (muxbus, sign-in required) exists for inter-agent messaging across hosts — everything else stays local.
Per-Agent Identity
Each agent runs under its own account, with API keys held in the OS keychain and swappable per agent.
Memory-Safe Backend
Rust backend with compile-time memory safety. No buffer overflows, no use-after-free, no GC pauses.
Detailed posture in the docs: Trust model, Network exposure, Data sovereignty.
Compliance Framework Coverage
Building toward comprehensive coverage across the frameworks that matter most to regulated enterprises and government agencies.
Design maps to the Govern, Map, Measure, and Manage functions
Built toward event logging, risk classification, human oversight, and transparency
AI management system certification pathway
Audit readiness planned
Local-first architecture provides inherent data residency
Cryptographic module validation for government use
Audit Trail
AgentMux writes a structured event log of agent activity to your machine. It runs in-process, stays local, and is fully auditable. Tamper-evident sealing and compliance-grade export are on the roadmap.
Logged Today
- Agent registration and lifecycle events
- Inter-agent messages, cryptographically signed and sender-verified — per-agent HMAC signing for same-machine delivery, Ed25519 signing for LAN, and Ed25519 signing for our own review-notification service over WAN
- Every message carries a trust marker: verified sender, a failed signature, or an unverified/self-declared identity — never presented as more certain than it is
- Agent identity and session timestamps
- Structured events, written in-process
- Broader prompt, tool-call, and file-operation logging in development
Log Characteristics
- Structured, machine-parseable events
- Stored locally, in your user profile
- Inspectable in plain text
- Tamper-evident sealing — in development
- General agent-to-agent WAN signing beyond reviewed-PR notifications — roadmap
- SIEM export and retention policies — roadmap
- Compliance-grade reporting — roadmap
What stays local vs. what leaves the machine: Data sovereignty.
Supply Chain Security
Open source means you can verify, not just trust. Read the code, check the dependencies, build it yourself.
Open Source
LiveApache 2.0. Full source available. Community-verified. No proprietary dependencies.
100% Rust Backend
LiveThe backend and sidecar are memory-safe Rust with compile-time guarantees — no GC or unsafe-memory vulnerabilities in that layer. The bundled Chromium (CEF) UI host is not written in Rust.
Dependency Auditing
Livecargo-audit integration for continuous vulnerability scanning of the dependency tree.
Code Signing
PartialmacOS builds are code-signed and notarized today. Windows and Linux signing is on the roadmap.
Reproducible Builds
RoadmapDeterministic build pipeline so anyone can verify the binary matches the source.
SBOM Generation
RoadmapSoftware Bill of Materials for every release. Full dependency transparency.
Update + tool-download model in the docs: Update model.
Deployment Models
From individual developer workstations to air-gapped environments, run AgentMux wherever your security requirements put it.
Desktop
Standard installation on employee workstations. Agents and their event logs stay on the machine.
- Windows, macOS, Linux
- Per-user agent configs
- Local event log
- Zero network dependency
Enterprise Managed
IT-deployed with shared policy presets, so agents launch from configurations your team controls.
- Group policy / MDM deployment
- Locked-down agent configs
- Centralized log collection (roadmap)
- RBAC enforcement (roadmap)
Air-Gapped
Fully offline operation for classified and restricted environments. No internet connection required.
- Zero internet dependency
- Offline installer package
- Local-only data path
- FISMA / CMMC ready
Posture per deployment topology: Trust model.
Identity & Access
Enterprise-grade identity management for AI agents and their operators. Coming in Phase 3 (Claw).
RBAC
RoadmapRole-based agent permissions. Control which agents each team member can launch, which tools agents can access, and what data they can touch.
Enterprise SSO
RoadmapSAML and OIDC integration for single sign-on. Every agent session tied to an authenticated enterprise identity.
PIV/CAC Support
RoadmapSmart card authentication for government and defense environments. Hardware-backed identity verification.
Session Attribution
RoadmapEvery agent action attributable to an authenticated user session. Full chain of custody from human to outcome.
Today's credential model: Identity & credential storage.
Ready to govern your AI agents?
Whether you're preparing for EU AI Act compliance, building FedRAMP-ready infrastructure, or eliminating shadow AI in your organization — we'd like to help.