Skip to content
Back to skills

Codegrid

ASecurity

CodeGrid is a native macOS canvas where multiple coding agents (Claude, Codex, Gemini, Cursor, Grok, shells) run side by side in panes and collaborate via a local agent bus — no tmux, no cloud, no account, no stored API keys. Install this skill when an agent should know how to operate inside a CodeGrid pane, drive the workspace from outside (control socket or codegrid:// deep links), spawn or message sibling agents, or coordinate multi-agent work (delegate, review, pipeline, parallel fan-out,...

  • 14 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added June 4, 2026
ai-agentsshellbashgitapi

Works with

  • cursor
  • api
  • mcp

Security analysis

A100/100

Pro scans all 3 files and shows the line behind each finding

Scanned June 4, 2026

npx -y skills add Eskyee/agentbot-opensource --skill codegrid --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Codegrid?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Codegrid
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/eskyee-codegrid/badge)](https://www.skillsdirectory.com/skills/eskyee-codegrid)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: codegrid
description: >-
  CodeGrid is a native macOS canvas where multiple coding agents (Claude, Codex,
  Gemini, Cursor, Grok, shells) run side by side in panes and collaborate via a
  local agent bus — no tmux, no cloud, no account, no stored API keys. Install
  this skill when an agent should know how to operate inside a CodeGrid pane,
  drive the workspace from outside (control socket or codegrid:// deep links),
  spawn or message sibling agents, or coordinate multi-agent work (delegate,
  review, pipeline, parallel fan-out, monitor, debate). The differentiator:
  multiple coding agents collaborating on one canvas, addressable by stable
  session_id, with a read → message → read protocol built for orchestration.
metadata:
  product: CodeGrid
  vendor: ZipLyne LLC
  homepage: https://codegrid.app
  docs: https://codegrid.app/docs
  token: $GRID (Base, 0x6B456E66524aEC1792013eF9DFE87e3F84311ba3)
  version: 1
---

# CodeGrid — agent operating manual

CodeGrid runs many coding agents side by side on an infinite 2D canvas. Each
pane is one process (`claude`, `codex`, `gemini`, `cursor-agent`, `grok`, or a
shell) with its own working directory, git branch, and stable `session_id`. The
agent bus lets any pane discover, read, and message any other pane locally.

This skill is the umbrella entry point. Two references cover the full surface:

- **[Operating CodeGrid](./references/using-codegrid.md)** — the mental model
  (canvas / pane / workspace / agent / bus), MCP tools (`list_agents`,
  `read_pane`, `message_agent`), the control-socket JSON-RPC API
  (`agent_list`, `agent_read`, `agent_send`, `open_folder`, `new_session`,
  `new_workspace`), the `codegrid://` deep-link scheme, and an operating
  playbook with recipes.

- **[Agent-bus collaboration](./references/codegrid-agent-bus.md)** — the
  `read → message → read` protocol, identifying agents by role and
  `session_id`, orchestration patterns (delegate, review, pipeline, parallel
  fan-out, monitor, debate), etiquette and scope safety, loop/runaway
  prevention, failure recovery, and worked end-to-end examples.

## When to load which reference

| You're about to… | Load |
|---|---|
| Discover what's running, open a folder, spawn a session, or drive CodeGrid from outside | `using-codegrid` |
| Hand work to / consult / review with another agent already running in CodeGrid | `codegrid-agent-bus` |
| Both | Load `using-codegrid` first, then `codegrid-agent-bus` |

## Quickstart (inside a CodeGrid pane)

```text
list_agents()                                  # discover panes (session_id, role, status)
read_pane(<session_id>)                        # see what they're doing — always safe
message_agent(<session_id>, "[from <you>] <self-contained request>")
… wait, then read_pane(<session_id>) for the reply.
```

## Quickstart (from outside CodeGrid, e.g. another tool)

```bash
SOCK="$(cat ~/.codegrid/socket-path 2>/dev/null || echo ~/.codegrid/socket)"
printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"agent_list"}' | nc -U "$SOCK"

# Or deep-link the installed app:
open "codegrid://open?path=/abs/path/to/repo&type=codex"
open "codegrid://new"
```

## Install

```text
install the codegrid skill from https://github.com/BankrBot/skills/tree/main/codegrid
```

## Source

- Canonical source of truth lives in the CodeGrid repo under `skills/`
  (`skills/using-codegrid/SKILL.md`, `skills/codegrid-agent-bus/SKILL.md`).
  This Bankr Skills provider is a published copy kept in sync.
- App download + full docs: <https://codegrid.app>.
- Token (Base): `0x6B456E66524aEC1792013eF9DFE87e3F84311ba3` —
  see <https://codegrid.app/token>.

## Safety defaults an agent should respect

- **Read before you write.** `read_pane` is free; never message a busy agent.
- **Address by `session_id`**, never pane number — pane numbers shift.
- **Be explicit about scope.** Other agents may be in YOLO/autonomous mode and
  will act on whatever you send. Say "propose only, do not edit files" when you
  want analysis.
- **One message, then wait.** Multiple messages interleave into the target's
  input box and corrupt it.
- **Bound debates and don't spawn unbounded helpers.** Converge and report to
  the user.
- **Stay local.** Everything is same-machine IPC; there is no remote bus.

Full guidance: see the two reference files.

Files in this skill

  • SKILL.md4.3 KB
  • references/codegrid-agent-bus.md9.9 KB
  • references/using-codegrid.md9 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.