Changelog

[0.5.0] - 2026-10-05

Agent orchestration and specialist runtime architecture

• Expanded the specialist catalog across architecture, backend/API/data, UX/UI, quality, security, performance, and operations.
• Preserved governance and specialist responsibilities as separate taxonomy layers.
• Kept orchestration as the core execution model: planning, delegation, execution, verification, review, and handoff.
• Preserved provider-neutral Agent × Model execution, recovery, audit identity, replay policy, and idempotency guarantees from the 0.4.x runtime line.
• Formalized platform-aware research and development specialist routing as the canonical agent taxonomy.
• Kept MCP and Tool capabilities as existing execution boundaries without expanding them as product architecture.
• Removed cost-free positioning from the core product architecture; deployment and provider cost remain operational concerns.

Verification

• Specialist catalog contract coverage added for the expanded development disciplines.
• Recovery, replay-policy, tool identity, and idempotency regression coverage retained.
• Release gate requires the full test suite plus package/install validation before publishing the 0.5.0 release.

[0.4.3] - 2026-10-03

Project-scoped local MCP and Secure MCP Tunnel lifecycle

• Isolated managed local MCP services by project path so multiple projects can run independently.
• Hardened macOS MCP install/uninstall lifecycle handling and service identity.
• Bound Secure MCP Tunnel launchd services to explicit project MCP and health endpoints.
• Added project-scoped macOS tunnel status and uninstall scripts.
• Allowed project-scoped macOS tunnels for arbitrary project roots.
• Added lifecycle tests for macOS MCP services and tunnel launchd scripts.
• Preserved the Windows local MCP launcher behavior while isolating platform-specific tests.

Verification

• Full test suite: 309 tests passed, 3 skipped on macOS.
• Verified managed MCP LaunchAgent remains running after macOS reboot/login.
• Verified project-scoped Secure MCP Tunnel launchd service remains running after reboot.
• Verified source distribution and wheel build successfully.
• Verified clean reinstall of the generated multiagentos-0.4.3 package and CLI startup.

[0.4.2] - 2026-10-02

Local MCP service

• Made local Streamable HTTP MCP the primary cost-free quick-start path.
• Documented the verified ChatGPT client boundary: ChatGPT Web can use GitHub and local MCP, while ChatGPT Mobile is verified for GitHub access only.
• Added multiagentos mcp install, status, and uninstall for OS-native local MCP lifecycle management.
• macOS uses per-user launchd; Windows uses per-user Task Scheduler.
• The managed service listens on loopback and starts automatically at user login.
• Removed the Responses API / Secure MCP Tunnel smoke test from the core release path.
• Kept Secure MCP Tunnel as an optional remote integration rather than a baseline requirement.

All notable changes to MultiAgentOS are documented here.

[0.4.1] - 2026-10-01

Packaging and release metadata

• Corrected the PyPI project description source to use the repository README.
• Removed stale project identity text from the published metadata path.
• Added canonical Homepage, Repository, Documentation, Changelog, and Download links.
• Prepared the release workflow for a fresh PyPI and GitHub release.
• Updated installation guidance for macOS/Linux users to use python3.

[0.4.0] - 2026-09-29

Cost-free development baseline

• Enforced the COSTFREE-001 acceptance path in GitHub Actions.
• Verified clean-install bootstrap without a provider API key.
• Verified Agent Execution Runtime MCP initialization and tool discovery.
• Verified filesystem READ/WRITE through the MCP runtime.
• Verified unified patch application and readback.
• Verified shell/process execution through the Agent Execution Runtime MCP surface.
• Kept provider SDKs optional; the core package has no provider runtime dependency.
• Documented Secure MCP Tunnel integration and the current ChatGPT Web plan boundary.

CI

• Added a dependency-free end-to-end acceptance harness.
• GitHub Actions now blocks the baseline when install, bootstrap, MCP, filesystem, patch, or process validation fails.

Release scope

This release establishes the first explicit cost-free-by-default development baseline for MultiAgentOS. Paid AI providers, ChatGPT integrations, and other model clients remain optional integration layers.

[0.3.0] - 2026-09-28

Added

• Provider-neutral model capability discovery and normalization.
• Native provider discovery defaults for OpenAI, Anthropic, and Gemini.
• Persistent capability, quota, and health intelligence.
• Unified model control-plane state and operational event inspection.
• Quota-aware and health-aware model routing with cooldown recovery.
• Runtime observation of model usage and capabilities.
• Provider-neutral routing explainability through AIRouter.explain().
• Agent Execution Runtime routing explanation API and multiagentos models explain CLI.
• Model discovery and intelligence CLI commands.

Compatibility

• Existing AIRouter.assign() behavior remains available.
• Provider-specific metadata is normalized before it reaches routing decisions.
• Provider credentials remain environment/configuration concerns and are not persisted by the runtime.

[0.2.0]

The previous development baseline. See Git history for the detailed implementation sequence.