oh-my-agent: shared MCP daemons and offline BGM rendering

We just merged a complete overhaul of how oh-my-agent manages MCP processes. Previously, every agent session spawned its own copy of every stdio MCP server. We measured 65 processes and 800MB of RAM across six sessions. Now, serena daemons are shared per-project by default.

What's new

• BGM rendering: The video agent now supports --music <preset> to render background music beds offline.
• Strudel integration: Music is generated via Strudel in a headless Chrome instance without adding npm dependencies to the CLI.
• Unified config: Source control management configurations are now consolidated into a single .agents/oma-config.yaml file.

The music presets are arranged rather than looped, using scale degrees and per-bar gain ramps to build a proper bed over the specified duration.

What's fixed

• Global install scope: Fixed a bug where oma install --global registered the home directory as a project, preventing the language detection scanner from walking up to 20,000 irrelevant files.
• Hook paths: The oma-hook.sh wrapper no longer bakes in the absolute path of the installer, keeping version-controlled hooks clean across teams.
• Model probing: Fixed the model:probe command to correctly resolve custom registered slugs instead of always falling back to the owner prefix.
• MCP bridge paths: Stopped the bridge from writing machine-specific absolute paths into the committed .mcp.json file.

These fixes prevent local installation environments from leaking into version control and breaking workflows when shared across different machines.

What's better

• Daemon lifecycle: The new bridge mode starts a single serena HTTP daemon per project and shares it across sessions. The daemon shuts down automatically after a 10-minute idle grace period.
• Artifact footprint: Removed 298 files and 4.8MB of duplicated skill artifacts from the repository by dropping the generated/ tracking directory.
• Prompt engineering: Dropped explicit self-verification instructions and output-shape examples from core skills to align with Anthropic's new context guidance.
• Skill length linting: Added a warning when a skill body exceeds the 500-line limit recommended by prompt engineering guides.

The prompt reductions were validated across 20 agent runs, proving that removing explicit examples actually improved the agent's ability to fix mutable sort keys without introducing false positives.

Installation

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.sh | bash

# Windows (PowerShell)
irm https://raw.githubusercontent.com/first-fluke/oh-my-agent/main/cli/install.ps1 | iex

oh-my-agent is built for teams who orchestrate more than they prompt. Next up: expanding the video compositor with custom branded layers.

https://github.com/first-fluke/oh-my-agent