multi-core

Reviewed at commit 46e1c7a New version, not yet reviewed

Caution: Routes all of Claude Code's API traffic through a local gateway, adds a PATH line to your shell profile, and lets some non-Claude models act unchecked in auto mode.

Sends data to The providers you pick: ChatGPT (chatgpt.com), OpenCode Zen (opencode.ai), Cursor (api2.cursor.sh), Antigravity, Grok

Claude Code plugin: use ChatGPT (Codex), Cursor, OpenCode Zen, and Antigravity models inside one Claude Code session. Switch with /model, delegate to named…

Made by
greenpolo
What we checked

Adds a launch command (claude-multi) that starts a gateway on your machine and routes Claude Code's requests through it, so you can pick ChatGPT, Cursor, OpenCode Zen, Antigravity or Grok models in /model and as subagents, using each provider's own login. Your conversation and workspace context go to the provider you select. In auto mode, Zen, Antigravity and Grok actions run without a safety check; plain claude sessions are unaffected.

Review details

Static review of the source at the pinned commit; not run. The diff since 6ee7ba3 (about 195k characters, 41 files) was too large for a delta review, so the whole plugin was reviewed again from the full bundle and the changes were read closely in the clone. The marketplace source is the repository root, so the whole repo is installed; hooks.json loads plugins/multi-core/hooks/register.ts. Hooks: tool.call (no matcher; answers only the mcp__multi-core__ display rows the gateway issued and passes every other call on), tool.check (allows a display row only with a gateway-issued token, denies any other call to those names), tool.describe, agent.offer, agent.spawn, classic.SubagentStart, classic.PreToolUse (Agent/Task only, adds quota advice as additional context when switched on in the usage pane), classic.UserPromptSubmit, classic.SessionStart, session.start, session.end, session.compact, turn.step, turn.complete, command.run (/multi-usage), ui.render, ui.message. Runtime calls: $.env.get, $.http.fetch (only to the loopback gateway URL from the environment, with a token and per-session key), $.state atoms, $.tool.register, $.command.register, $.ui.status/toast/open/invalidate, $.clock. Every hook returns early unless MULTI_MOD_GATEWAY_URL and MULTI_GATEWAY_TOKEN are set, which happens only under the launcher; then they post session id, cwd, permission mode, model and tool names to the gateway, resolve or deny provider subagent spawns, and send compaction transcripts to the gateway for harness models. The user-invoked setup skill (allowed-tools: Bash, disable-model-invocation) runs node src/setup.ts, which writes ~/.local/share/multi-cli and appends a marked PATH block to ~/.bashrc, ~/.zshrc, fish config or the PowerShell profile; status and uninstall skills run the installed local helper; no npx or package installs. The launcher binds an HTTP gateway on 127.0.0.1 with a random token, sets ANTHROPIC_BASE_URL to it and starts claude with temporary --settings and --agents; it also runs claude, codex app-server, agy and grok CLIs. Claude requests are forwarded with your own headers to api.anthropic.com. OpenAI requests reuse the tokens in ~/.codex/auth.json against chatgpt.com/backend-api/codex; Zen uses opencode.ai/zen; Cursor usage uses api2.cursor.sh. With Antigravity enabled, each launch installs a global PreToolUse hook in ~/.gemini/config/hooks.json (multi uninstall removes it); agy runs with --dangerously-skip-permissions under that hook's allow/deny policy translated from your Claude settings, and the hook script now decides locally without a network call. Changes in this version: the gateway now answers Claude Code's auto-mode safety check itself for every tool call in a non-Claude reply (gateway/verdicts.ts). Read-only and coordination tools pass, plan mode blocks file writes, and for OpenAI each other action goes to the codex-auto-review model on chatgpt.com with your autoMode and permission rules and up to 768 KiB of the request history, plus an inspect_path tool that reads files or lists folders inside the working directory; denial, failure or a 60 s timeout blocks the action, with at most four reviews at once. Zen, Antigravity, Grok, and OpenAI when no reviewer was found at launch get an automatic not_flagged verdict in default, acceptEdits, auto, dontAsk and bypassPermissions modes; the status line says "actions run unreviewed". Previously the launcher disabled auto mode when the OpenAI reviewer was missing and Claude was not signed in; that setting is removed. Cursor keeps its own native review, which fails open. Also removed: the classifier-parsing approval bridge, tool observer and /multi/permission route; worker model selection moved from a tool.call hook into the gateway; malformed plugin agent frontmatter is reparsed leniently and refuses only native workers of that type; a signed-out agy or grok no longer stops the launch; old harness records are migrated on load. The diff adds no new hosts, processes or file paths. The hook modules, setup, the prompt files, the launcher changes and every changed gateway and provider file (verdicts.ts, approval.ts, review-evidence.ts, review-limit.ts, provider.ts, harness-session.ts, harness-migration.ts, agent-definitions.ts, mod-routes.ts, server.ts serveReply) were read fully; the rest of the gateway and provider folders were checked by targeted search of every fetch, host, process, file write and permission decision. npm dependencies (@cursor/sdk, yaml, js-tiktoken) and the external CLIs were not reviewed. Network use matches the repository's README, PRIVACY.md and CHANGELOG. No eval or obfuscation found.

What it uses
  • Reads your prompts: yes
  • Sees tool calls: yes
  • Blocks or changes tool calls: yes
  • Reads or writes files: yes
  • Runs programs on your machine: yes
  • Calls the model (uses tokens): no
  • Sends data over the network: yes
  • Stores data: yes
  • Shows something on screen: yes
Hooked events

Claude Code plugin: use ChatGPT (Codex), Cursor, OpenCode Zen, and Antigravity models inside one Claude Code session. Switch with /model, delegate to named native workers, keep each provider's own login and permissions.

Made this mod? Add a badge to your README

modsforClaude badge for multi-core

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/multi-core.svg)](https://modsforclaude.com/mods/multi-core/)

← All mods