I Built A Local Control Plane For My Coding Agents
Using different AI agents used to be a mess.
I used Copilot, Claude Code, and OpenCode. Each tool had its own home directory, its own settings, and its own memory. Every time I added a new skill or rule, I had to ask: where does this tool expect it?
My home directory became a junk drawer of AI folders.
I tried using Markdown files to pass plans between agents. It helped, but it failed. One agent would miss an assumption, and the next agent would produce the wrong result. Files provided data, but they did not provide shared context.
I needed a way to:
- Write high-level rules once.
- Keep permission rules in one place.
- Stop copying skills by hand.
- Share memory across different projects and sessions.
I built a local control plane to solve this.
The core of my setup is a single directory: ~/.ai.
Instead of scattered folders, I use symlinks. The tools think they are in their own space, but the real structure lives in one place. I route all model traffic through a local proxy and use a shared SQLite database for memory via MCP.
This separates two different jobs:
- Routing: Model requests go through a local proxy before reaching the provider.
- Memory: Long-lived facts live in a local store.
If a model behaves strangely, I check the proxy. If an agent forgets a detail, I check the memory. This makes debugging easy.
My shared memory stores things that are expensive to rediscover:
- Project decisions.
- Architecture notes.
- Workflow preferences.
- Recurring bugs.
I also centralized my security. I use shared files for coding standards and denied commands. This ensures that switching from one agent to another does not change how safe the environment is.
This is not about finding one perfect agent. It is about making sure my agents share enough context so that switching between them does not feel like starting over.
Source: https://dev.to/sperekrestova/i-built-a-local-control-plane-for-my-coding-agents-1bf7
Optional learning community: https://t.me/GyaanSetuAi