Skip to content

Codex skill/MCP: message Grok Bot via gbot send #25

Closed

Description

@ScriptedAlchemy

Summary

Ship a Codex-facing skill (and optionally a thin MCP/plugin) so Codex agents can reach Grok Bot the same way we already reach bots from the shell: gbot send <bot-or-group> <message>.

Parent: track under the bidirectional bridge epic.

Why this first

Reverse direction (Codex→Grok) has no app-server attach problem. gbot already talks to the Grok Bot gateway. Teaching Codex to call it unblocks coordination immediately.

Proposed shape

  • Skill (or plugin skill) with clear description: when to ping Grok, how to name the bot/group, message format
  • Optional MCP tool wrappers: gbot_send, gbot_thread (read recent chat)
  • Document auth on Codex hosts: GROK_BOT_GATEWAY_URL + GROK_BOT_GATEWAY_TOKEN, Grok Bot app session, or CURSOR_ACCESS_TOKEN
  • Example AGENTS.md snippet for repos that want default routing to General or a specialist bot

Acceptance criteria

  • Installable from this repo (Codex plugin marketplace entry or documented skill path)
  • gbot send works from a Codex turn with sandbox/network policy called out
  • README section: “Talking to Grok Bot from Codex”
  • Smoke-tested on ubuntu-main and Mac Studio Codex CLI / Desktop where gbot is on PATH

Non-goals

  • Replicating Desktop’s “Sent by ChatGPT from another task” UX
  • Scraping Desktop IPC pipes

Activity

  1. ScriptedAlchemy commented on Sep 15, 2026

    @ScriptedAlchemy
    OwnerAuthor

    GPT-6 Pro packaging guidance (chatgpt.com, 2026-09-14)

    Ordering

    1. Skill first — when to send, target selection, submission vs completion.
    2. Thin MCP over the same code: gbot_send, gbot_thread (+ optional restricted target list).
    3. Plugin packages skill + MCP config only — no separate business logic.

    Implementation corrections

    • MCP should import existing modules (commands.js / gateway), not shell out to gbot send and parse stdout.
    • Do not expose bot deletion, arbitrary gateway methods, or a general shell tool.
    • Fix CLI message/flag separation (--stdin, --) so agent text cannot be eaten as flags.
    • Expose gateway clientNonce / replyToId via CLI if useful for correlation.
    • Use current OpenAI packaging: root plugin.json, skills/, mcp.json (.codex-plugin/ as fallback only).
    • Ensure npm files allowlist includes skills/plugin/docs or publish will omit them.

    Suggested tool profile

    gbot mcp --tools grok
    

    Acceptance still stands; add:

    • MCP calls modules in-process (no subprocess parse)
    • npm pack / clean install includes skill + MCP assets
    • Safe stdin/-- message path smoke-tested

    Source: model 6 Pro on chatgpt.com.

  2. added a commit that references this issue on Sep 15, 2026
    181d710
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions