MCP for coding tools
Let an AI coding tool — Claude Code, Cursor or Codex — read your workspace's
runs, devices and cloud data. The phloem CLI adds Phloem's MCP server and
your sign-in to the tool's config.
Set up
Install the CLI, sign in, then name your tool:
curl -fsSL https://dl.phloem.ai/cli/install.sh | shirm https://dl.phloem.ai/cli/install.ps1 | iexphloem login
phloem mcp setup claude-code # or: cursor · codex| Tool | Where the entry goes |
|---|---|
claude-code | added with Claude Code's claude mcp add-json, user scope |
cursor | ~/.cursor/mcp.json, beside your other servers |
codex | ~/.codex/config.toml, replacing the phloem table in place |
Add --print to show the entry instead of writing it, for another tool or to
add it by hand. Restart the tool to see Phloem's tools.
What the tools can do
| Reads | Changes (asks first) |
|---|---|
| list test runs, filtered by device, recipe, unit, outcome or project | send a device its configuration now |
| run a read-only SQL query on your workspace's cloud data | store a device's configuration and apply it |
| list run fields, and show one | |
| list devices — name, fleet, online | |
| list profiles, and where each runs | |
| show a device's configuration — fleet, device, merged, last applied |
Every tool acts as you and sees only what you can see. A tool that changes
something asks you first. If the AI tool cannot ask, it refuses and points you
to Computer or the phloem CLI. Applying a configuration is refused while a
setup is deployed on that device. No tool drives hardware.
The server runs in the cloud at your workspace's address; there is
no local one. On a computer that is offline or not signed in, an AI tool uses
the phloem CLI directly instead.
Updated