Skip to main content

Sync skills, MCP servers, commands and agents across your tools

Each AI tool keeps its own skills, MCP servers, commands and agents. This guide copies them between the tools on one machine, and installs your MCP servers on another machine.

chat-recall writes to these tools: Claude Code (claude), Codex (codex), Antigravity (agy), Cursor (cursor), OpenCode (opencode), Hermes Agent (hermes) and pi (pi). The name in brackets is the tool id that the commands below take.

Copy everything to every tool on this machine​

  1. Run the sync:

    chat-recall toolkit sync

  2. Read the report. It names each artifact it copied, and to which tool.

The sync copies each skill, MCP server, command and agent to every tool that does not have it yet. It never overwrites an artifact that a tool already has.

To copy only some types, give --types. The types are skill, mcp, command and agent:

chat-recall toolkit sync --types skill,mcp

Commands and agents change format on the way. For example, a Claude Code command in Markdown becomes a TOML command for Antigravity, and a Claude Code agent becomes a TOML agent for Codex. Instruction files get the name each tool reads: CLAUDE.md, AGENTS.md or GEMINI.md.

Copy one artifact​

chat-recall toolkit copy <type> <name> <fromTool> <toTool>

For example, to copy the skill pdf from Claude Code to Codex:

chat-recall toolkit copy skill pdf claude codex

Install your MCP servers on a new machine​

  1. Install chat-recall on the new machine and log in:

    chat-recall login

  2. See what the pull will install. This writes nothing:

    chat-recall toolkit pull --dry-run

  3. Install the MCP servers:

    chat-recall toolkit pull

  4. If the report names environment variables, set them for each MCP server that needs them.

The server keeps an inventory of your toolkit. For an MCP server, the inventory holds the full registration: the command, the arguments, the URL, the allowed tools and the names of its environment variables. The values of the variables never leave the machine they are on, so step 4 is necessary when a server needs a key.

The server does not hold the files of skills, agents and commands. The pull reports them by name, and you copy them yourself or share them with chat-recall team publish (see the CLI reference).

An MCP server installs by itself only on a machine with the same operating system as the device that registered it, so a macOS-only server does not reach a Linux machine. To install one anyway, name it:

chat-recall toolkit pull --mcp acme-mcp,example-mcp

Sync from the web app or from an assistant​

The Toolkit view of the web app shows which artifact exists on which tool and on which device. A sync that you start there waits in a queue, because only the target machine can write its own config files. The target machine runs the queue the next time its background service runs, or when you run:

chat-recall toolkit drain

An assistant does the same with two MCP tools:

  • recall_toolkit_status shows which artifact exists on which tool and device.
  • recall_toolkit_sync queues the work. With scope: "pull" it installs the account's MCP servers on a device. With scope: "fan_out" it copies one device's artifacts between that device's tools.

Install the chat-recall skills​

The chat-recall skills tell an agent when to use each recall_* tool. chat-recall init installs them. To install them again, or into a tool that you added later:

chat-recall install-skills

--all-tools installs them into every supported tool, also the tools that are not found on this machine. --uninstall removes them.