ZCode in NeuroSquadZCode,with a canvas around it
Run as many ZCode agents as you need, each the real zcode in its own card. ZCode’s own hooks report every turn and every question, arrows hand it terminals, notes and other agents, and each card resumes its session and counts every request it makes.
Official sitezcode.z.ai- The real
zcode, your own providers - Your
~/.zcodeis only read - Free, for Windows and macOS
It knows when ZCode is done — and when it’s asking
A quiet terminal can mean “finished” or “waiting for your permission”. NeuroSquad doesn’t guess: each card’s settings carry ZCode’s own hooks, and ZCode waits for every one of them.
UserPromptSubmithookYou press Enter. The card’s icon spins, the Squad status card counts it as working, and the toast from its last question is taken down.
PermissionRequesthookZCode asks before a call. The toast names it — ZCode needs your permission: Bash rm -rf dist — and stays until you answer; the tool’s PostToolUse puts the card back to working.
StophookThe turn is over. The next queued prompt goes out, the journal gets the answer, and a toast says it’s done.
Questions and plans count too
ZCode sends AskUserQuestion and ExitPlanMode through the same PermissionRequest, so the toast reads ZCode is asking: Tabs or spaces? and the card waits on you.
Esc is read from the key
Escape cancels a turn (“Turn cancelled.”) without any hook, so the card settles from the key itself. Ctrl+C on an idle ZCode quits it — the card never sends it.
Subagents don’t move it
A subagent runs in a session of its own (sess_subagent_…); its hooks never change the card’s status.
Three things you’ll do with ZCode
Replicas of the app, played step by step. When ZCode asks, the answer is yours.
ZCode fixes a failing test through a terminal card, stops on its permission dialog before deleting a folder, and finishes — the journal writes itself.
Everything ZCode can do in NeuroSquad
Measured on ZCode runtime 0.16.9 through the zcode command of zcode-app-cli. Most of it is ZCode’s own — hooks, MCP, sessions, its usage ledger. Where NeuroSquad had to build something, or it can’t be done, the card says so.
- 01
Knows its state
ZCode’s own hooks report every turn, permission and question, so nothing is guessed.
Finished and waiting are different
Each card’s settings carry ZCode’s
Docs: Finished and waiting are differentprocesshooks —SessionStart,UserPromptSubmit,PermissionRequest,Stop— and ZCode awaits each one, so the card learns of every turn as it happens.A notification that says what it needs
The toast names the call —
Docs: A notification that says what it needsZCode needs your permission: Write notes.md— or the question ZCode asks. One per agent; a waiting one stays until you answer.- Built by NeuroSquad
Esc, never Ctrl+C
Escape cancels a running turn; ZCode sends no hook for it, so the card settles from the key. Ctrl+C on an idle ZCode quits it, so the budget brake and the phone’s key row never send it.
Docs: Esc, never Ctrl+C Sessions that survive a restart
ZCode names its session on the first prompt and has no flag to choose one;
Docs: Sessions that survive a restartSessionStartreports it, so the next launch iszcode --resume <id>. A deleted session starts fresh by itself.The Squad status card
Every AI agent of the workspace on one card, with its status, model and turn time.
Docs: The Squad status card
- 02
Works through the canvas
An arrow from the card is a set of tools, over ZCode’s own MCP client.
Arrows are tools
The card’s settings add NeuroSquad’s MCP server with the token as a placeholder. An arrow to a terminal, note, browser or agent gives ZCode those tools.
Docs: Arrows are toolsNeuroSquad’s own tools never prompt
Every
Docs: NeuroSquad’s own tools never promptmcp__neurosquad__*tool is in the card’spermission.allowedTools, so they run without a dialog. Everything else still asks the way ZCode always does.- Not possible
No live tool list
ZCode reads a server’s tools once and never again. So a card gets every NeuroSquad capability from the start, and the arrows decide at call time.
Docs: No live tool list Canvas mode that holds
Docs: Canvas mode that holds--disallowedToolsremoves Bash, its background shells and the web tools from what the model sees — yolo included — and aPreToolUsehook refuses them besides.MCP servers from the registry
An installed server comes as a second server with no pre-approval: ZCode shows its own dialog before every call to it.
Docs: MCP servers from the registry
- 03
Built for long runs
Context, Compact, a prompt queue and a journal — handled on the card.
Context meter
The latest request’s tokens against the model’s window, from ZCode’s own rules for that model. The ring turns amber from 80%.
Docs: Context meterCompact
The card’s Compact button sends ZCode’s own
Docs: Compact/compact; its call shows up in Usage too.Prompt queue
Line up the next messages: each goes out after
Docs: Prompt queueStop, as one paste — short, multi-line or long.Journal and hand-off
The last answer is read from the card’s own session database: into a connected note, into a fresh agent’s first prompt, back to a lead waiting in
Docs: Journal and hand-offagent_wait.Subagents, counted apart
ZCode’s
Docs: Subagents, counted apartAgenttool runs a subagent in a child session. Its requests are counted to the card and marked as a subagent’s in Usage, Run Stats and Agent Pulse.
- 04
Works in a team
Every card has a ZCode folder of its own, so any number of agents share a project without sharing a session list.
- Built by NeuroSquad
A ZCode folder per card
ZCode reads its settings from your home folder and has no switch for it, so a small preload points ZCode — and only ZCode — at the card’s copy: your settings plus NeuroSquad’s. Secrets stay placeholders on disk.
Docs: A ZCode folder per card - Built by NeuroSquad
Instructions
ZCode doesn’t read an MCP server’s instructions, so the card’s
Docs: InstructionsSessionStarthook hands them over as extra context. Your ownAGENTS.mdis read as always. Isolated worktrees
Give an agent its own git worktree, so parallel agents never edit the same files. The card’s ZCode folder follows the card.
Docs: Isolated worktreesDangerous mode, per card
Docs: Dangerous mode, per card--mode yolo: no permission dialogs. Switching it restarts the card with the same session; switching it off resumes in your usual mode. Canvas mode still holds.Model and CLI switch
Pick the model per card, let a lead agent change it with
Docs: Model and CLI switchagent_set_model, or move the conversation to Claude Code and back — the session comes along.
- 05
Cost, control and access
What it spends, what it may spend, which model it runs on, and how you reach it.
Exact usage, every request
Read from ZCode’s own usage ledger: one row per request, titles, compaction and subagents included, cache reads and writes split out. A model with no known price shows “no price”, never $0.
Docs: Exact usage, every requestBudget brake
Past the workspace limit each ZCode gets Escape — the turn stops, ZCode keeps running — and automatic prompts wait until you raise the limit.
Docs: Budget brakeOpenRouter and your own servers
A provider of the card’s own for OpenRouter, with NeuroSquad’s attribution headers, or for your own server (Chat Completions or Anthropic Messages). The key comes from the card’s environment, never from disk.
Docs: OpenRouter and your own serversFrom your phone
Scan a QR code and the whole canvas opens in the phone’s browser: read ZCode’s terminal, answer its dialog, send the next prompt.
Docs: From your phoneNot in WSL or over SSH yet
The preload and the Node launcher would have to live inside the target, so ZCode cards run on this computer only for now.
Docs: Not in WSL or over SSH yet
Who’s waiting, and what it cost
Two cards that answer the questions you ask most when several ZCode agents run at once.
The Squad status card
Every agent of the workspace with its status, model and turn time. The ones waiting on you come first, with what ZCode asks. Tap a status to filter.
- coupon-fixZCodeGLM-5.3ZCode needs your permission: Bash rm -rf distNeeds you2:34
- api-migrationZCodeGLM-5.3Move routes to the new clientWorking14:58
- lint-setupZCodeGLM-5.3-FlashZCode is asking: Tabs or spaces?Needs you3:51
- docs-passZCodez-ai/glm-4.6Update the READMEFinished7 min. ago
Usage & cost
Read from the usage ledger in each card’s own ZCode database, so every request belongs to exactly one card. Tokens and cost add up to the total exactly.
| Agent | Tokens | Cost | ||
|---|---|---|---|---|
| coupon-fix | 1,184,310 | $0.96 | ||
| api-migration | 802,554 | $0.64 | ||
| docs-pass | 196,402 | $0.11 | ||
| Total | 101 | 2,183,266 | $1.71 |
The real ZCode, in a folder of its own
NeuroSquad starts the same zcode you run yourself, in a real terminal in the card. This is everything it adds:
Your ~/.zcode is only read
The preload works inside ZCode’s own processes only: your settings file is read, the card’s copy is written. A marker in a user’s setting.json survived every test run byte for byte.
No secrets on disk
Tokens and keys are ${NAME} placeholders in the card’s files, filled in memory and removed from ZCode’s environment — so its shell commands never see them.
Which ZCode
Z.ai publishes ZCode’s source; the zcode command on npm comes from zcode-app-cli, which ships Z.ai’s runtime unchanged. The setup wizard installs that one; Z.ai’s own build runs the same way.
ZCode in NeuroSquad, answered
Give ZCode a canvas
Free for Windows and macOS. Your code, your providers and your costs stay on your machine.