4. AI: agents and local models
4.2 Shared instructions, skills and MCP servers
A single folder to instruct all the agents at once: AGENTS.md, the skills, the kit's MCP servers and those of a project.
In the app: «AI › Agents»
In short
- Instructions, skills and MCP servers are written only once, in the data\agents folder, and apply to all the agents.
- The kit syncs them when the app starts, before and after each agent and when an agent ends.
- You can also edit an agent's files: the change passes to the others.
- The kit's MCP servers can also be added from the AI page, “Shared by the agents” row, “The kit's MCP servers” box, with the ready servers GitHub, Playwright and Fetch one click away.
- MCP servers for a single project are added in the project's “Editor” section.
What is shared
| What | What it is for | The file to edit |
|---|---|---|
| Global instructions | Rules and preferences that apply in every project (“reply in English”, “use pnpm”, your coding style). | data\agents\AGENTS.md |
| Skills | Reusable procedures that the agent loads when needed, one folder per skill with a SKILL.md file. | data\agents\skills\<name>\SKILL.md |
| MCP servers | Programs or services that give the agent new tools (reading documentation, querying a service, using the browser). | data\agents\mcp.json |
- On the AI page, choose the “Shared by the agents” row and press “Open shared folder” (it is also on the Agents pane's title).
- Edit AGENTS.md with a text editor, add folders in skills or servers in mcp.json.
- The change reaches the agents at their next start (or right away, if you restart the app).
data\agents\mcp.json{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
},
"docs": {
"type": "http",
"url": "https://example.com/mcp"
}
}
}The kit's MCP servers and the ready servers
- On the AI page, choose the “Shared by the agents” row: below “Instructions and skills” there is the “The kit's MCP servers” box, the list of data\agents\mcp.json.
- Under “Ready servers” press “Add” beside the server you need; for another server press “Add a server” and fill in name, type and program or address.
- Atollo writes mcp.json and syncs the agents at once: Claude Code, Codex, Antigravity and Pi have the server without waiting for their next start. “Remove” on a row removes it.
| Ready server | What it does | Where |
|---|---|---|
github | Issues, pull requests and repositories on GitHub, with the kit's GitHub account. It is GitHub's own server, online (api.githubcopilot.com/mcp), reached through Atollo, which keeps the token: nothing to download. | the kit |
playwright | A browser the agent opens and uses: pages, clicks, forms, screenshots. | the kit |
fetch | Reads a web page and gives it to the agent as text. | the kit |
filesystem | Reads and writes files, only in the project's folder. | the project |
mysql | Queries the project's MySQL database (DBHub), once it has been created. | the project |
mongodb | Queries the project's MongoDB database, once it has been created. | the project |
How the sync works
- Instructions: AGENTS.md is copied as CLAUDE.md for Claude Code and as AGENTS.md for Codex, Antigravity and Pi. If you edit one of those files, the change goes back into AGENTS.md and on to the others. If both sides change, the most recent wins and the other is kept in the conflicts folder.
- Skills: each agent sees the skills folder through a link (copies on FAT32 and exFAT drives). A skill created inside an agent is moved to data\agents\skills and becomes everyone's; a skill deleted in an agent is set aside and can be recovered.
- MCP servers: mcp.json is translated into each agent's format. With Claude Code it works in both directions: a server added with claude mcp add -s user goes back into mcp.json. Codex, Antigravity and Pi receive the kit's servers and leave alone the ones you added by hand in their configurations. Pi does not accept SSE servers: the kit leaves them out of its mcp.json, with a warning.
MCP servers for a single project
- Open the project, “Editor” section, “Project MCP servers” box.
- For a ready server press “Add” beside it under “Ready servers”: filesystem (the files, only in the project's folder) and, once the project's database has been created with “Create database and user”, mysql or mongodb. For another server go on with the next step.
- Press “Add a server”.
- Write the “Name” (letters, digits, - and _) and choose the “Type”: “program” (with “Program”, for example npx, and the “Arguments, one per line”) or “address (http)” (with the “Address”).
- Press “Add”. “Remove” on a row removes it.
These servers apply only in that project, in addition to the kit's. Claude Code reads them from the project's .mcp.json file (and asks for confirmation on first use); Codex from .codex\config.toml, which Atollo keeps in sync; Antigravity and Pi use only the kit's servers. The programs they need (Node for npx, uv for uvx) are installed when the agent starts; the project's database reaches the mysql or mongodb server as a variable, never written in .mcp.json, which may be in the repository.
Under the hood
| Original | Copies for the agents |
|---|---|
data\agents\AGENTS.md | data\home\.claude\CLAUDE.md · data\home\.codex\AGENTS.md · data\home\.gemini\AGENTS.md · data\home\.pi\agent\AGENTS.md |
data\agents\skills | data\home\.claude\skills · data\home\.codex\skills · data\home\.gemini\antigravity-cli\skills · data\home\.pi\agent\skills |
data\agents\mcp.json | data\home\.claude\.claude.json · data\home\.codex\config.toml · data\home\.gemini\config\mcp_config.json · data\home\.pi\agent\mcp.json |
- The sync state, the conflicts and the deleted skills are in data\agents\.sync (conflicts, deleted).
- Sync warnings end up in the app's log, data\app\logs\app.log.
- You can also run the sync from the command line:
Atollo.exe sync [--project <folder>] [--verbose]. - Claude Code, Codex and Pi trust a project's folder only after you open an agent in that project from Atollo: from then on they can run commands there without asking for trust every time. The other folders in projects are not trusted in advance.
- In Codex ${VAR} variables are not expanded: the kit translates them where Codex has a dedicated key, otherwise it copies them as they are with a comment.
- Pi keeps everything in data\home\.pi\agent (the PI_CODING_AGENT_DIR variable). Besides instructions, skills and MCP servers, the kit writes there models.json (the “atollo” provider: the local models on disk, with the server's port and key, owner-only) and settings.json (on Windows the kit's Git Bash as the shell, project trust “ask” if you have not set it, the local model as the default until you choose another provider). Your own providers and settings stay.