Appearance

MyChatBot plugin for Claude Code and Codex ​

The MyChatBot plugin connects Claude Code and Codex to the Sales Platform, Agents Platform, content, and public Docs operations through one MyChatBot connection. Its workflow skills help Claude Code or Codex choose the owning platform, inspect the relevant account configuration, use the live MCP schema, and verify what changed.

The plugin does not force every request through one end-to-end build process. It keeps read-only inspection, customer-data access, saved configuration, private tests or billed previews, activation, customer contact or publication, and destructive actions as distinct stages. Approval for one stage does not imply approval for the next.

The plugin is distributed from the public MyChatBot source repository. You can install it directly in either host; an Anthropic or OpenAI directory listing is not required.

Live account

The plugin connects to live account state. Installation and browser authorization do not authorize an account change. Review the exact targets and effects of each write, test, activation, customer contact, publication, replacement, or deletion stage separately.

What the plugin covers ​

Ask in ordinary language. You do not need a slash-command prefix. Common tasks include:

  • inspect subscription and usage capacity, Sales assistants, knowledge and integrations, channels, Agents inventory, and relevant routine readiness, then report blockers, duplicates, unverified areas, and a dependency-ordered setup plan;
  • create or tune customer-facing Sales assistants for sales, support, booking, ordering, website chat, voice, or messaging;
  • configure Agents Platform agents, skills, connectors, Business Knowledge, routines, schedules, and triggers;
  • connect a website as stored or live knowledge, and add or update FAQs, manual product catalogs, product feeds, and Agents Business Knowledge attachments;
  • inspect and configure Sales channels and integrations or Agents connectors, with OAuth and secret-entry steps handed back to the user;
  • inspect or update pipelines, labels, leads, notes, tasks, orders, calendars, and conversation records within an approved record scope;
  • rehearse Sales assistant conversations privately and run repeatable eval scenarios before customer-facing activation;
  • validate routines, stage disabled schedules and triggers, run tool-free previews, and activate approved automations;
  • prepare disabled follow-ups, scheduled messages, outbound calls, or outreach campaigns, then contact customers only after exact approval;
  • generate media or speech, inspect social analytics and ads, and publish or schedule social posts only after separate publication approval.

The plugin selects the relevant workflow skill, routes to the owning platform, reads fresh state, and uses the current MCP input schema. Skills teach process and judgment; MyChatBot services remain authoritative for tenancy, schemas, plan gates, validation, and writes.

One connection and browser signup ​

Both hosts configure one MCP resource:

https://api.mychatbot.app/api/mcp/plugin

It composes the existing Sales, Agents, content, and public Docs tool implementations in the API. It does not route tool calls through the browser Claude connector or expose a reduced connector catalog.

When the host starts OAuth, connector.mychatbot.app serves only as the authorization and signup service. The browser asks for an email address and a one-time code. It reconnects the account for that verified email or creates a MyChatBot account when none exists. The user does not visit the dashboard first or copy an mcp_… key into the host.

Install in Claude Code from source ​

Run:

bash
claude plugin marketplace add https://github.com/mychatbot-app/mychatbot-agent-plugin.git#main
claude plugin install mychatbot@mychatbot-app

Then start browser sign-in directly from an interactive terminal:

bash
claude mcp login plugin:mychatbot:mychatbot

Complete the browser flow, wait for success, and start a new Claude Code conversation so the authenticated tools are loaded.

When Claude itself is running setup through a non-interactive Bash tool, it can launch the same command in a pseudo-terminal without relying on a plugin path:

bash
login_log="$(mktemp "${TMPDIR:-/tmp}/mychatbot-login.XXXXXX.log")"
python3 -c 'import pty,sys; pty.spawn(sys.argv[1:])' \
  claude mcp login plugin:mychatbot:mychatbot > "$login_log" 2>&1 &
printf 'MyChatBot sign-in started. Follow progress in %s\n' "$login_log"

The unique log is only for progress messages. Do not print credential material from it, and do not start concurrent login attempts.

On Windows, run the claude mcp login command directly from an interactive PowerShell window.

Install in Codex from source ​

Run:

bash
codex plugin marketplace add mychatbot-app/mychatbot-agent-plugin --ref main
codex plugin add mychatbot@mychatbot-app

Codex requests authorization during installation. If it still needs sign-in, run:

bash
codex mcp login mychatbot

Complete the same browser flow and start a new Codex thread.

Verify the one MCP resource ​

For Claude Code:

bash
claude plugin details mychatbot@mychatbot-app
claude mcp get plugin:mychatbot:mychatbot

For Codex:

bash
codex mcp get mychatbot

The server URL must be:

https://api.mychatbot.app/api/mcp/plugin

If the server exists but no MyChatBot tools appear, authenticate again rather than reinstalling. Tool lists are captured at session start, so open a new conversation after a successful sign-in.

Complete tool inventory ​

The one connection exposes all 172 bundled operations:

Owning platformToolsMain jobs
Sales Platform117Sales assistants, CRM, knowledge, integrations, channels, outreach, orders, calendar, private tests, and evals
Agents Platform35Agents, skills, connectors, Business Knowledge, routine authoring, readiness, previews, schedules, and triggers
Content17Media and speech generation, social posting, analytics, and ads
Docs3get_docs_structure, search_docs, and read_docs_page over public documentation

The plugin does not copy the full schemas into its prompts. The host uses the live tool description and JSON Schema. A checked operation map intentionally assigns all 172 bundled operations and the ten on-demand Product operations to workflow skills, while the separate risk contract preserves approval classes.

Product MCP remains per catalog ​

The Product MCP has ten read-only search and catalog-inspection tools, but each catalog has a different credential-bearing URL. A universal plugin must not embed or guess that value.

For ordinary setup, configure the catalog through Sales, reuse it in Agents as Business Knowledge, and verify customer behavior through private Sales tests or evals. When direct semantic catalog queries are required, open that catalog's Connect AI tools flow and add its Product MCP as a separate read-only server. See Product MCP.

First read-only turn ​

The plugin starts with the smallest relevant inventory:

  • Sales or a cross-platform audit: get_account_summary, followed by targeted list or detail reads;
  • Agents work: get_account_authoring_inventory;
  • routine work: get_routine_authoring_context, its returned Markdown references, get_account_authoring_inventory, and existing routines;
  • content work: only the models, accounts, task state, or reports required by the explicit request.

A failed or partial inventory is unknown, not empty. A general audit does not read leads, chats, messages, individual orders or bookings, audience previews, exports, eval transcripts, or routine-session history. When the task needs customer data, the plugin first bounds the records, fields, and date range, then minimizes identifying detail in its report.

Approval and verification model ​

The plugin separates:

  1. read-only account inspection;
  2. customer-data reads;
  3. reversible configuration;
  4. customer-record changes;
  5. private tests;
  6. billed generation or preview;
  7. activation;
  8. customer communication or publication;
  9. disabling, full replacement, cancellation, disconnection, reset, or deletion.

Approval for one class does not authorize the next. The plugin resolves identifiers from fresh reads, shows complete values before full replacement, executes an approved write once, and re-reads persisted state. It investigates an ambiguous result instead of automatically retrying a write.

Agents routines move from authoring context and inventory through validation, complete YAML review, save, readiness, and an optional separately approved billed, tool-free preview. Schedules and triggers are created disabled and enabled only after their own approval.

Three easily missed boundaries are explicit in version 0.4.0:

  • schedule_message requires the exact recipient, sender, content, timezone, and time to be approved before scheduling;
  • update_integration_trigger first runs with confirm=false, shows the exact before and after state, then requires approval before confirm=true;
  • get_routine_session_history may expose customer inputs, agent outputs, and tool details, so it is used only for a bounded, explicitly requested diagnostic.

Update or troubleshoot ​

Refresh the source marketplace:

bash
claude plugin marketplace update mychatbot-app
codex plugin marketplace upgrade mychatbot-app

Update or reinstall the plugin, authenticate if requested, and start a new conversation. If authentication fails, do not paste credentials into chat or a command. Restart the browser login and complete its email verification flow.

Agent-readable version ​

AI hosts can read this page directly as Markdown at:

https://docs.mychatbot.app/agents/claude-code-plugin.md

See also ​

Copyright © 2023–2026 MyChatBot