ZCode

ZCode is Z.ai’s coding agent built around its GLM models. Z.ai published its source under Apache-2.0; the zcode command runs the same agent runtime as the ZCode desktop app, in your terminal. In NeuroSquad each ZCode agent is a card with the real zcode running inside.

Measured on ZCode runtime 0.16.9 (zcode-app-cli 3.14.4-32).

Install

Install the zcode command

Z.ai publishes no ready-made terminal download, so the zcode command comes from npm:


npm install -g zcode-app-cli@latest

zcode-app-cli is an unofficial client that ships Z.ai’s runtime unchanged. Settings → Setup in NeuroSquad can install it for you. A zcode you built from Z.ai’s own repository works too.

Set up a provider

Set up ZCode’s providers and default model the way you normally do — cards start from your providers. A card can also run on OpenRouter or on your own server instead.

Add a ZCode card

Add an agent card and choose ZCode. NeuroSquad checks that the zcode it finds really is ZCode — unrelated tools with the same name are not used.

What works

Exact status. ZCode’s own hooks tell the card when it is working, finished, or waiting for you — with the call it wants to make (ZCode needs your permission: Bash rm -rf dist) or the question it asks. See Finished / needs your input.
Arrows. NeuroSquad’s MCP server is in the card’s settings, and its own tools never prompt. ZCode reads a server’s tools once, so the card gets every NeuroSquad tool from the start and the arrows decide at call time. A skill or an installed MCP server needs the card’s next start.
Canvas mode. Bash, its background shells and the web tools are removed from what the model sees, even in dangerous mode. See Canvas mode.
Sessions. After a restart the card continues the same conversation (zcode --resume). A deleted session is replaced by a fresh one.
Long sessions. Context meter, Compact (ZCode’s /compact), the prompt queue, the journal and hand-off.
Usage. Every request is read from ZCode’s own usage ledger — titles, compaction and subagents included, subagents marked as such — in Usage & costs, Run Stats and Agent Pulse.
Models and providers. Pick the model per card; a lead agent can change it with agent_set_model. OpenRouter and your own servers work, with the key kept out of every file.
Switching CLIs. Move the conversation to Claude Code and back.
Plugins. Memory, Context7, Caveman and Graphify reach ZCode’s turns; the token saver works in dangerous mode.
From your phone. The card works in remote access like any other.

Dangerous mode

ZCode cards start in dangerous mode with --mode yolo. Switching it on or off restarts the card on the same conversation; turning it off resumes in your usual mode. See Dangerous mode.

Good to know

Your ~/.zcode is only read. Each card has its own copy of your settings plus NeuroSquad’s, its own session database and, when needed, its own provider file. Keys and tokens are never written to disk.
Stop a turn with Esc. Ctrl+C on an idle ZCode quits it, so NeuroSquad never sends it — the budget brake and the phone’s key row use Escape.
ZCode’s project memory is off in a card unless your own setting.json turns it on. Its background calls are recorded nowhere, so Usage couldn’t count them; if you turn it on, those calls are missing from Usage.
Conversations live in the card’s database. A zcode --resume in your own terminal finds them only with ZCODE_SESSION_DB_PATH pointing at that file. Deleting a card keeps its database.
After Escape, ZCode’s terminal UI swallows the next key you press.

ZCode cards run on this computer only — not yet in WSL or SSH workspaces.