CortexLink

One plugin. Every CLI. Stay in flow.

You know the feeling. You are deep in a vibe coding session — ideas flowing, your agent keeping up — and then you need to hand something off to another CLI tool. Suddenly you are out of the zone, hunting for the right flags, checking docs, figuring out what --allow-tool even accepts. The flow breaks.

CortexLink fixes that.

CortexLink is a cross-agent orchestration plugin for your AI agent CLI. It turns your active AI agent into a persistent control center that decomposes tasks, delegates to other CLI agents, and manages the full workflow — while you stay focused on what you're building.

Note: CortexLink is designed for use with 2 or more AI CLI platforms. To get the most out of cross-agent workflows, you need access to at least two supported CLI agents (e.g. Copilot CLI + Claude Code).

How it works

CortexLink uses a Control Center → Agents model:

CONTROL CENTER (your AI agent — decomposes, routes, reviews)
  ├── AGENT (Copilot CLI)
  ├── AGENT (Claude CLI)
  └── AGENT (future-cli ...)
        └── [own tools & sub-agents — internal, platform-native]

• The control center is your active AI agent. It holds the plan, fans out tasks to the right CLI agents (in parallel or in sequence), and reviews every report before proceeding.
• Each agent is a full CLI agent in its own right — not just an executor. It can use its own platform-native tools and sub-agents internally, then self-verifies and reports back.
• The tree is one level deep. Agents are peers — they don't chain to each other. Width scales as you add agents; depth stays fixed.
• Every agent returns a structured report (~150 words: status, steps, files changed, issues)

No context bloat. No manual coordination. The workflow just happens.

Why CortexLink

Stay in flow. Your active CLI knows how to route tasks, which agent to use, and how to handle the results — without you managing it.

Reduce your token spend. Every delegated subtask runs in a separate subprocess. Only the final result comes back into your context — no reasoning chains, no intermediate tool calls.

Prerequisites

• Claude Code — CortexLink is a Claude Code plugin
• At least one supported CLI agent installed and authenticated (see Get started below)

Get started

Before installing, make sure you have at least one supported CLI agent set up.

GitHub Copilot CLI:

brew install copilot-cli
copilot login

Claude CLI:

npm install -g @anthropic-ai/claude-code
claude auth login

Then install CortexLink from inside Claude Code:

/plugin marketplace add namos2502/agent-plugins
/plugin install cortexlink@agent-plugins
/reload-plugins

Warning

If commands don't appear after /reload-plugins, fully quit Claude Code (Cmd+Q on macOS, close the window on Windows) and relaunch. /reload-plugins does not always pick up new commands without a full restart.

Then run setup once:

/cortexlink:setup

This detects which CLI agents are installed, verifies authentication, and activates CortexLink — the SessionStart hook will inject orchestration context automatically on every future session.

Updating

To get the latest version:

/plugin marketplace update agent-plugins
/plugin update cortexlink@agent-plugins

Then restart Claude Code to pick up any new or changed commands.

Uninstalling

To remove CortexLink completely:

1. Run /cortexlink:cleanup — removes the CortexLink sections added to ~/.claude/CLAUDE.md and ~/.copilot/copilot-instructions.md
2. Run /plugin uninstall cortexlink@agent-plugins — removes the plugin files

Troubleshooting

Commands don't appear after /reload-plugins Fully quit Claude Code (Cmd+Q on macOS, close the window on Windows) and relaunch. /reload-plugins does not always pick up new commands without a full restart.

The SessionStart hook doesn't seem to be injecting context Run /cortexlink:doctor — it checks the hook script, verifies jq is installed, and test-runs the hook to confirm it produces valid output.

Setup reports an agent as not found but it's installed The agent binary may not be on Claude Code's PATH. Open a terminal and run which copilot or which claude. If it's found there but not in setup, your shell PATH may differ from Claude Code's environment — check your shell config (~/.zshrc, ~/.bashrc).

Auth check fails but I'm already logged in Re-run the login command (copilot login or claude auth login). Plugin sandboxing can sometimes make cached tokens appear invalid — a fresh login usually fixes it.

Delegation fails or the agent doesn't respond Run /cortexlink:doctor to check agent auth and hook status. Verify the agent binary is on PATH (see above). Make sure the agent isn't already running interactively in another terminal.

Command	What it does
/cortexlink:setup	One-time setup — detects agents, authenticates, registers CortexLink as always-on in ~/.claude/CLAUDE.md
/cortexlink:doctor	Diagnose the SessionStart hook, plugin files, and agent auth — asks about symptoms before running checks
/cortexlink:cleanup	Remove CortexLink configuration added by setup
/cortexlink:help	Show plugin information and commands

Under the hood

CortexLink uses a skill-per-concern architecture:

skills/
  orchestration/
    SKILL.md          ← control center protocol, routing decision tree, report format
  agents/
    copilot-cli/
      SKILL.md        ← Copilot CLI behavioral reference
    claude-cli/
      SKILL.md        ← Claude CLI behavioral reference
    [future-agent]/   ← drop here to add a new agent
      SKILL.md

After /cortexlink:setup, the SessionStart hook injects the orchestration protocol automatically on every new session — no extra command needed. Run /cortexlink:doctor if anything seems off.

Routing

The orchestration skill teaches your CLI when and how to delegate:

Task type	Route to
GitHub-specific (PRs, repos, Actions)	Copilot CLI
General code tasks, Anthropic-native reasoning	Claude CLI
Needs current session context or open files	Handle natively
Host has a native subagent for it	Use it (faster, no auth needed)

Supported agents

CLI	Cmd	Status
Copilot CLI	copilot -p	✅ Supported
Claude Code	claude -p	✅ Supported

License

This project is licensed under the MIT License