Your external agent

Connect Claude, Codex, Copilot, Cursor or another AI assistant to Nexa Agora, choose what it may do, and pause or remove it.

For everyoneConnect your tools · page 7 of 35

Your external agent lets you work on all of Nexa Agora with the AI assistant you already use, such as Claude, Codex, GitHub Copilot or Cursor. The assistant works here as you: it has exactly your permissions, never more, and every action it takes is recorded under your name. It thinks on the assistant's own plan, so the workspace's credits stay for Nexa's agents. This page helps you connect an assistant, check what it can do, and narrow, pause or remove it.

Before you start

Every member can connect their own assistants. Each connection is personal.
Open Governance › Settings, then Your external agent in the Settings index.
Have your assistant open on the computer or app where you use it.

Settings › Your external agent for Sara Lin, with no assistant connected yet and the Connect an assistant button. — 01Connect the AI you already use, and it works here as you.

Pick your assistant

Click Connect an assistant.
The panel Connect an external assistant asks Which assistant do you use?
Pick yours: Claude, Claude Code, ChatGPT, Codex, GitHub Copilot, Cursor, Antigravity, Grok Build, Mistral Vibe, Hermes Agent, ChatGPT · GPT builder for a Custom GPT, or Another MCP client.

The "Which assistant do you use?" panel with every assistant tile, and the "Not ready yet — and why" fold opened. — 02Pick the assistant you use. The fold below lists the ones that cannot connect yet, and why.

The fold Not ready yet — and why lists Mistral Vibe Work, Cursor cloud agents, Gemini Enterprise and the Gemini app, each with its reason.

Connect an assistant

A guided sheet, Connect followed by your assistant's name, walks you through five steps: Where and plan, How it connects, What it may do, Connect and First call.

Under Where do you use …?, pick where you use the assistant, for example the web, the desktop app or your code editor.
Under Which … plan are you on?, pick your plan with that assistant.
The page tells you whether that plan can connect.
Under How should it connect?, choose one way:
A key you paste once (Recommended) — the page shows a key once and fills it into the steps for you. You decide what the assistant may do before the key exists.
Sign in with your Nexa account
no key to copy. The assistant opens a Nexa page and you sign in.

Some assistants, and some places you use them, offer only one way. With Claude, the code editor (Claude Code) offers both; the web and the desktop app offer one each.

The Claude Code connect sheet at "How should it connect?", with "A key you paste once — Recommended" and "Sign in with your Nexa account". — 03A key lets you choose what the assistant may do before it connects.

With a key, choose under What may it do?:
Read and write
it works here as you: posts, plans, tasks and everything else your role allows.
Read only
it can look around, search and summarize. Every write is refused.
Only some activities
tick exactly what it may touch.

Optionally, type a name in Name this connection. Only you see it.

The "What may it do?" step with Only some activities chosen and a few activities ticked. — 04Choose exactly what the assistant may touch.

Create the key and copy it. Then click I've pasted it.
The steps for your assistant appear, with your key and Agora's address already filled in, each with a Copy button. Follow them in your assistant.
With sign-in instead, follow the steps, sign in with your Nexa account on the page that opens, approve, then click I've signed in.
At First call, the sheet shows Waiting for …'s first call… and an example question. Ask your assistant something about your workspace, for example "What happened in the Oceanic Teas project since yesterday?"
When the first call arrives, the sheet says the assistant is in. Click See what it can do.

The First call step showing "Claude is in." and See what it can do. — 05The connection is ready at its first call.

The connection is saved under Connected.

With Sign in with your Nexa account, the connection starts able to read and write everything your role allows, with deleting off. You narrow it on its page afterwards.

In Claude, open Settings › Connectors and choose Add custom connector.
Name it and paste Agora's address from the page into Remote server URL.
When Claude asks how to sign in, pick Use your own OAuth client. Paste the Client ID the page shows, and leave the secret empty.
On the Nexa page that opens, sign in with your Nexa account and approve.
Back in Claude, ask it something about your workspace.

Connect ChatGPT on the web

ChatGPT connects with Sign in with your Nexa account. Your ChatGPT plan decides whether it can: Plus connects; on Business, Enterprise and Edu, a ChatGPT workspace admin first allows custom MCP connectors in Workspace settings › Permissions & roles. Free can't add one.

In ChatGPT on the web, open chatgpt.com/plugins, choose +, then Add custom MCP server.
Name it, for example Nexa Agora, and paste Agora's address from the page into Server URL. The icon and the description are optional.
Leave Authentication on OAuth. Nothing else to fill in: ChatGPT tells Nexa who it is by itself.
Tick I understand and want to continue and click Create as a plugin.
On the Nexa page that opens, sign in with your Nexa account. If you belong to more than one workspace, choose the one ChatGPT will work in, then approve.
Back in ChatGPT, ask it something about your workspace. The connection appears under Connected at its first call.

Connect a Custom GPT in ChatGPT

Choose ChatGPT · GPT builder.
The sheet Connect a Custom GPT opens.
Type Its name in Agora. What the GPT posts in Agora is signed with this name. Optionally, type What it is for.
Under What may it do?, choose Read and write or Read only. Click Create its credentials.
In ChatGPT's GPT builder, open your GPT, then Configure › Actions › Create new action:
In Authentication, choose OAuth and paste the Client ID, Client Secret, Authorization URL, Token URL and Scope from the page. Leave Token Exchange Method on Default (POST request).
In Schema, choose Import from URL and paste the schema address from the page.
Save.
Click I've saved the action.
Copy the Callback URL that OpenAI shows under the action. Paste it into The callback URL from OpenAI, then click Save and finish.

Without the callback address, the GPT cannot sign in. If you click Add it later, the connection reads Set-up incomplete — one address missing until you add it.

Connect another MCP client

Choose Another MCP client.
The sheet Connect another MCP client asks Which client?
In your client, open its custom connector settings and add a new one.
Fill in the fields the page shows: Client ID, Authorization URL and Token URL. Leave Client Secret empty; there is no secret. Set Token Auth Method to none (PKCE).
As the server type, choose Streamable HTTP and paste the address from the page.
Save it, and sign in with your Nexa account when the client asks.
The connection appears under Connected.

See what a connection can do

Under Connected, click Open on the connection.
Its page shows Acts as you, with your name and role, and It can never do more than you can.
Read the panel What … can do in Agora. It groups activities:
Reads
your assistant does the thinking; Agora is the place it reads.
Writes
still your work; it lands under your name, through your assistant.
Hands work to a colleague
a Nexa agent runs on its own model. This spends the workspace's credits.
Admin territory
only for workspace admins.
Use the filters Everything, Can do now and Off or needs more.

A connection's page with "What Claude can do in Agora" and the Reads, Writes, Hands work to a colleague and Admin territory groups, with a "Needs a workspace admin" badge highlighted. — 06Each activity says whether the assistant can do it, and if not, why.

Each activity is marked, for example Can do, Off · read-only, Off for this connection, Off · needs your consent or Needs a workspace admin.

Narrow, pause, re-key or remove a connection

On the connection's page, open Settings.
Every change holds from the connection's next call.
Change what you need:
Name
rename the connection, then click Save.
Active
turn it off to pause the connection. Every call is refused; its settings and key stay as they are. Switch back on resumes it.
Can also write
turn it off to allow reading and searching only.
Can also delete
off by default, so deleting and what cannot be undone stay with you.
Only chosen activities
tick the activities that stay on. Allow everything my role allows resets it.
Replace key
stops the old key at once and gives you a new one, shown once.
To end the connection, click Delete for a key connection or Disconnect for a sign-in connection, then confirm with Yes, delete it or Yes, disconnect.

The connection's Settings with Active and Can also write on, Can also delete off, Only chosen activities, Replace key and Delete. — 07Narrow or pause a connection at any time.

For a sign-in connection, also remove the connector in your assistant's own settings, so it stops asking.

What you will see

The Connected list, with Assistant, Type, Connects with, Status and Permissions. Statuses include Active, Paused, Read-only, Can write, Limited · N and Set-up incomplete.

The Connected list for Sara Lin: Claude, Active and Can write; Claude Code, Read-only; and a paused Cursor connection. — 08The Connected list shows each assistant's status and permissions.

Paused. — the assistant gets a clear refusal on every call.

Custom GPTs are closing on a Custom GPT's page: OpenAI stops every GPT on 11 December 2026 and no longer lets personal accounts create new ones. Your GPT keeps working here until then. Its status reads Closing 11 Dec.

Signing in with Nexa isn't set up for … on this server yet, so these steps can't work. — go back and choose a key instead.

Couldn't load your assistants.

Good to know

The assistant's own thinking runs on its own plan. Only work it hands to a Nexa colleague, such as chatting with a named agent, launching a plan or generating in Design Studio, spends the workspace's credits.
What an assistant may do is your role, narrowed by the connection's choices. Your choices can narrow your role, never widen it.
Deleting agents, people, knowledge or memory, invitations, Brain rules or a project's status is off by default and needs its own switch.
Some things are always yours alone: passkeys, two-step sign-in and account recovery, admin tokens, handing the workspace to a new owner, payment methods, and deleting the whole workspace. No assistant does these, whatever it is allowed.
Every connection also appears in Settings › Security under Signed in, labeled External agent with its name.
Sign-in completes only when the client returns to a local address on your own computer (localhost or 127.0.0.1), or to a callback Nexa already has registered for that client, such as Claude's or ChatGPT's. A client hosted on the web with its own callback, or one that returns through its own app link, cannot finish it.
Claude Desktop with a key needs Node.js installed. Without it, the connection fails quietly.
Grok works through Grok Build.
A Custom GPT reaches only what is in the actions pasted into it: at most 30, OpenAI's limit.

Your external agent

Connect Claude, Codex, Copilot, Cursor or another AI assistant to Nexa Agora, choose what it may do, and pause or remove it.

For everyoneConnect your tools · page 7 of 35

Before you start

Every member can connect their own assistants. Each connection is personal.
Open Governance › Settings, then Your external agent in the Settings index.
Have your assistant open on the computer or app where you use it.

Pick your assistant

Click Connect an assistant.
The panel Connect an external assistant asks Which assistant do you use?
Pick yours: Claude, Claude Code, ChatGPT, Codex, GitHub Copilot, Cursor, Antigravity, Grok Build, Mistral Vibe, Hermes Agent, ChatGPT · GPT builder for a Custom GPT, or Another MCP client.

The fold Not ready yet — and why lists Mistral Vibe Work, Cursor cloud agents, Gemini Enterprise and the Gemini app, each with its reason.

Connect an assistant

A guided sheet, Connect followed by your assistant's name, walks you through five steps: Where and plan, How it connects, What it may do, Connect and First call.

Under Where do you use …?, pick where you use the assistant, for example the web, the desktop app or your code editor.
Under Which … plan are you on?, pick your plan with that assistant.
The page tells you whether that plan can connect.
Under How should it connect?, choose one way:
A key you paste once (Recommended) — the page shows a key once and fills it into the steps for you. You decide what the assistant may do before the key exists.
Sign in with your Nexa account
no key to copy. The assistant opens a Nexa page and you sign in.

Some assistants, and some places you use them, offer only one way. With Claude, the code editor (Claude Code) offers both; the web and the desktop app offer one each.

With a key, choose under What may it do?:
Read and write
it works here as you: posts, plans, tasks and everything else your role allows.
Read only
it can look around, search and summarize. Every write is refused.
Only some activities
tick exactly what it may touch.

Optionally, type a name in Name this connection. Only you see it.

Create the key and copy it. Then click I've pasted it.
The steps for your assistant appear, with your key and Agora's address already filled in, each with a Copy button. Follow them in your assistant.
With sign-in instead, follow the steps, sign in with your Nexa account on the page that opens, approve, then click I've signed in.
At First call, the sheet shows Waiting for …'s first call… and an example question. Ask your assistant something about your workspace, for example "What happened in the Oceanic Teas project since yesterday?"
When the first call arrives, the sheet says the assistant is in. Click See what it can do.

The connection is saved under Connected.

With Sign in with your Nexa account, the connection starts able to read and write everything your role allows, with deleting off. You narrow it on its page afterwards.

In Claude, open Settings › Connectors and choose Add custom connector.
Name it and paste Agora's address from the page into Remote server URL.
When Claude asks how to sign in, pick Use your own OAuth client. Paste the Client ID the page shows, and leave the secret empty.
On the Nexa page that opens, sign in with your Nexa account and approve.
Back in Claude, ask it something about your workspace.

Connect ChatGPT on the web

In ChatGPT on the web, open chatgpt.com/plugins, choose +, then Add custom MCP server.
Name it, for example Nexa Agora, and paste Agora's address from the page into Server URL. The icon and the description are optional.
Leave Authentication on OAuth. Nothing else to fill in: ChatGPT tells Nexa who it is by itself.
Tick I understand and want to continue and click Create as a plugin.
On the Nexa page that opens, sign in with your Nexa account. If you belong to more than one workspace, choose the one ChatGPT will work in, then approve.
Back in ChatGPT, ask it something about your workspace. The connection appears under Connected at its first call.

Connect a Custom GPT in ChatGPT

Choose ChatGPT · GPT builder.
The sheet Connect a Custom GPT opens.
Type Its name in Agora. What the GPT posts in Agora is signed with this name. Optionally, type What it is for.
Under What may it do?, choose Read and write or Read only. Click Create its credentials.
In ChatGPT's GPT builder, open your GPT, then Configure › Actions › Create new action:
In Authentication, choose OAuth and paste the Client ID, Client Secret, Authorization URL, Token URL and Scope from the page. Leave Token Exchange Method on Default (POST request).
In Schema, choose Import from URL and paste the schema address from the page.
Save.
Click I've saved the action.
Copy the Callback URL that OpenAI shows under the action. Paste it into The callback URL from OpenAI, then click Save and finish.

Without the callback address, the GPT cannot sign in. If you click Add it later, the connection reads Set-up incomplete — one address missing until you add it.

Connect another MCP client

Choose Another MCP client.
The sheet Connect another MCP client asks Which client?
In your client, open its custom connector settings and add a new one.
Fill in the fields the page shows: Client ID, Authorization URL and Token URL. Leave Client Secret empty; there is no secret. Set Token Auth Method to none (PKCE).
As the server type, choose Streamable HTTP and paste the address from the page.
Save it, and sign in with your Nexa account when the client asks.
The connection appears under Connected.

See what a connection can do

Under Connected, click Open on the connection.
Its page shows Acts as you, with your name and role, and It can never do more than you can.
Read the panel What … can do in Agora. It groups activities:
Reads
your assistant does the thinking; Agora is the place it reads.
Writes
still your work; it lands under your name, through your assistant.
Hands work to a colleague
a Nexa agent runs on its own model. This spends the workspace's credits.
Admin territory
only for workspace admins.
Use the filters Everything, Can do now and Off or needs more.

Each activity is marked, for example Can do, Off · read-only, Off for this connection, Off · needs your consent or Needs a workspace admin.

Narrow, pause, re-key or remove a connection

On the connection's page, open Settings.
Every change holds from the connection's next call.
Change what you need:
Name
rename the connection, then click Save.
Active
turn it off to pause the connection. Every call is refused; its settings and key stay as they are. Switch back on resumes it.
Can also write
turn it off to allow reading and searching only.
Can also delete
off by default, so deleting and what cannot be undone stay with you.
Only chosen activities
tick the activities that stay on. Allow everything my role allows resets it.
Replace key
stops the old key at once and gives you a new one, shown once.
To end the connection, click Delete for a key connection or Disconnect for a sign-in connection, then confirm with Yes, delete it or Yes, disconnect.

For a sign-in connection, also remove the connector in your assistant's own settings, so it stops asking.

What you will see

Paused. — the assistant gets a clear refusal on every call.

Signing in with Nexa isn't set up for … on this server yet, so these steps can't work. — go back and choose a key instead.

Couldn't load your assistants.

Good to know

The assistant's own thinking runs on its own plan. Only work it hands to a Nexa colleague, such as chatting with a named agent, launching a plan or generating in Design Studio, spends the workspace's credits.
What an assistant may do is your role, narrowed by the connection's choices. Your choices can narrow your role, never widen it.
Deleting agents, people, knowledge or memory, invitations, Brain rules or a project's status is off by default and needs its own switch.
Some things are always yours alone: passkeys, two-step sign-in and account recovery, admin tokens, handing the workspace to a new owner, payment methods, and deleting the whole workspace. No assistant does these, whatever it is allowed.
Every connection also appears in Settings › Security under Signed in, labeled External agent with its name.
Sign-in completes only when the client returns to a local address on your own computer (localhost or 127.0.0.1), or to a callback Nexa already has registered for that client, such as Claude's or ChatGPT's. A client hosted on the web with its own callback, or one that returns through its own app link, cannot finish it.
Claude Desktop with a key needs Node.js installed. Without it, the connection fails quietly.
Grok works through Grok Build.
A Custom GPT reaches only what is in the actions pasted into it: at most 30, OpenAI's limit.