Connect ChatGPT
Add Gorila to ChatGPT as a custom MCP server and keep it on with a custom instruction.
Steps
-
Open ChatGPT's plugins and click Add → Add custom MCP server. Open ChatGPT's plugins
-
Fill in the fields. Keep Authentication on OAuth, tick “I understand and want to continue” and save:
- Name:
Gorila - Server URL:
https://app.gorila.la/api/mcp
- Name:
-
Sign in with your Gorila account when ChatGPT asks. Allow it when asked.
-
Paste this block into Settings → Personalization → Custom instructions (or a project's instructions) and save.
Block to paste (Custom instructions):
your organization uses Gorila (MCP server "gorila") as its shared work memory: the team's tasks and company knowledge (brand tone, copy rules, SOPs, policies, decisions). ALWAYS: - Before writing anything for your organization (emails, customer replies, copy, ads, posts, docs) or running a recurring process, call get_relevant_knowledge with a short description of the situation, and follow what it returns. - Facts about your organization itself (product, pricing, customers, policies) are in Gorila: call get_relevant_knowledge before searching the web or leaving a placeholder. - When I correct your tone, a process or a fact that could apply again, call propose_knowledge_change. Nothing is published until a person approves it. - When I mention work for someone else ("tell Ana to…", "this is for Esteban"), offer to create_task with full context (background, goal, done criteria, links) so their AI can do it. - When I start working or ask what's next, call list_my_tasks. - When you start working on a Gorila task, set it to doing with update_task; when it's finished (for code: its pull request merged), set it to done with a comment of what you did. Don't wait for me to ask. - Put a Gorila task's id in the title or branch of its pull requests and comment_task the link; comment again when it's merged or closed. - Store distilled, reusable text in Gorila, never raw documents, full emails or chat logs. - If a person or a task is ambiguous, ask me. Never guess. -
Open a new chat (Custom instructions only apply to new chats) and try the prompt below.
-
Let your AI work between sessions: a nightly run turns what you committed to into tasks, prepares the most urgent ones and checks the rest, and never sends anything. Ask it: "let my AI work for me". It creates the run itself. With the plugin on "Allow low-risk", a run stops at one of its writes on ChatGPT's own approval and waits for you there. On "Allow all tools" it went through without stopping in our test, with the elevated risk ChatGPT warns about.
-
Optional, for the gorila as its icon: ChatGPT only sets a plugin's icon when it's added, so if you want it, remove the plugin, add it again and pick this PNG under Icon.
https://app.gorila.la/brand/gorila-icon-256.png
This step is required. Without it ChatGPT almost never checks Gorila on its own: it did in 2 of 10 test chats, and in 10 of 10 with the block.
Let your AI do it
Paste this into Claude in Chrome or Cowork and it follows the steps for you. Sign in to Gorila when ChatGPT asks, then open a new chat.
Set up Gorila in my ChatGPT account, in my browser:
1. Open https://chatgpt.com/plugins → Add → Add custom MCP server. Name it Gorila, use this Server URL, keep Authentication on OAuth, tick "I understand and want to continue" and save: https://app.gorila.la/api/mcp
2. When it asks me to sign in to Gorila, stop and let me do it.
3. Open Settings → Personalization → Custom instructions and add the text below, keeping what's already there. Save.
your organization uses Gorila (MCP server "gorila") as its shared work memory: the team's tasks and company knowledge (brand tone, copy rules, SOPs, policies, decisions).
ALWAYS:
- Before writing anything for your organization (emails, customer replies, copy, ads, posts, docs) or running a recurring process, call get_relevant_knowledge with a short description of the situation, and follow what it returns.
- Facts about your organization itself (product, pricing, customers, policies) are in Gorila: call get_relevant_knowledge before searching the web or leaving a placeholder.
- When I correct your tone, a process or a fact that could apply again, call propose_knowledge_change. Nothing is published until a person approves it.
- When I mention work for someone else ("tell Ana to…", "this is for Esteban"), offer to create_task with full context (background, goal, done criteria, links) so their AI can do it.
- When I start working or ask what's next, call list_my_tasks.
- When you start working on a Gorila task, set it to doing with update_task; when it's finished (for code: its pull request merged), set it to done with a comment of what you did. Don't wait for me to ask.
- Put a Gorila task's id in the title or branch of its pull requests and comment_task the link; comment again when it's merged or closed.
- Store distilled, reusable text in Gorila, never raw documents, full emails or chat logs.
- If a person or a task is ambiguous, ask me. Never guess.
- When a task's end comes after this session (a reply, a merge, an approval), set what it's waiting for and when to look again (waiting_on, recheck_at), not only a comment.