How I manage skills and MCP servers across AI coding agents

Keeping AI Coding Agents in Sync

Working with AI coding agents recently has meant switching between them
more often than I expected.

Sometimes a model hits token limits.
Sometimes rate limits kick in.
Sometimes a different model simply performs better for a task.

So during the same project I often move between tools like Codex,
Cursor, Gemini, Claude Code, or Copilot.

---

Most coding agents now support roughly the similar concepts:

• skills
• MCP servers
• local tools and config
• filesystem discovery

But they do not share state.
That leads to some practical problems.

• if I install a new MCP in one agent, I have to install it again in the other agents.
• If I update a skill, I have to make sure every agent is using the updated version.
• If I try a new agent, the entire setup has to be recreated again.

Over time the setups drift apart.

---

The Idea

I wanted one workspace to define the environment, with agents acting as
consumers of that definition.

That became skills-sync.

skills-sync is a CLI for keeping skills and MCP configuration in one workspace and syncing that state into supported agents.

Instead of treating every agent as a separate manual setup, it gives you a single source of truth.

How It Works

With skills-sync, your environment lives in a local workspace.

Tt uses symlinks so agents point to the same underlying files instead of keeping duplicated copies around.

That means:

• skills live in one place
• MCP configuration lives in one place
• updates propagate cleanly
• new agents can reuse the same setup
• switching tools no longer means rebuilding everything

Today the project supports Codex, Cursor, Gemini, Copilot and Claude Code.

Basic Workflow

A realistic first run looks like this:

# initialize the local workspace with starter content
skills-sync init --seed

# use the default profile
skills-sync use personal

# materialize runtime output for supported agents
skills-sync sync

Adding skills from an upstream is straightforward:

skills-sync profile add-upstream --source matlab/skills
skills-sync list upstream-content --upstream matlab_skills
skills-sync profile add-skill --upstream matlab_skills --path skills/matlab-test-generator

Once multiple agents are installed locally, you can also check whether
they have drifted from the workspace definition:

skills-sync agents inventory
skills-sync agents drift --dry-run

And if you want to move the same setup between machines, you can export
and import a profile instead of rebuilding it from memory.

Conclusion

skills-sync is still early, but the core workflow is already working well for me.

If you are juggling multiple coding agents and you are tired of their environments drifting apart, take a look:

https://github.com/ryanreh99/skills-sync
https://www.npmjs.com/package/@ryanreh99/skills-sync

Feedback or ideas are welcome.
Thanks for reading.

Comments

[Hamza KONTE]

Great topic — MCP server management is genuinely complex at scale. One thing I'd add: the quality of prompts you're giving each agent matters as much as which servers you're connecting. Poorly-scoped agent instructions compound across your stack.

I built flompt (flompt.dev) to help with exactly this — it's itself an MCP server you can add to any agent: claude mcp add flompt https://flompt.dev/mcp/. Agents can call decompose_prompt/compile_prompt to structure their own instructions. Free, open-source.