Use Chat.co experts in Codex

1. Availability

Codex

Manual setup

Manual configuration

Codex connects to the server through its own MCP configuration. There is no separate Codex directory listing.

Server URL: https://mcp.chat.co/mcp
Permission: chatco:read
Status reviewed: 30 September 2026

Codex adds Chat.co as a remote MCP server you configure yourself. We tested the full flow natively in Codex: listing experts, a cited answer, a follow-up in the same conversation, a direct search and opening a source. Chat.co has no separate Codex directory listing or approval.

2. Before you start

Your own Chat.co account. You sign in yourself during setup. Never use another person's login.
A workspace with at least one expert you can chat with in Chat.co, on an active plan. If someone else manages the account, they give you access to experts; see Team, roles & seats.
Message credits in that workspace for expert answers and standalone searches. Listing experts and opening sources are free.
Optional: permission to search an expert's knowledge directly. You only need it for standalone search. Asking questions and opening an answer's citations do not require it.
Codex installed and signed in, as described in OpenAI's Codex documentation. The commands below use the Codex CLI.

3. Add Chat.co to Codex

Add the server

bash

codex mcp add chatco --url https://mcp.chat.co/mcp

2
Sign in
bash
```
codex mcp login chatco
```
Your browser opens the Chat.co approval page. Approve it, then return to Codex.
3
Check it
codex mcp list shows chatco. Inside a Codex session, /mcp lists its four tools: list_authorized_experts, consult_expert, search and fetch.

Or edit config.toml

The same server can be declared directly in ~/.codex/config.toml. Add the block below, then run codex mcp login chatco:

~/.codex/config.tomltoml

[mcp_servers.chatco]
url = "https://mcp.chat.co/mcp"

In the Codex IDE extension and the Codex app, you can also add the server from Settings → MCP servers, using the same URL.

4. Approve in Chat.co

Every host sends you to the same Chat.co approval page. Sign in if asked, then check each part before you approve:

On the approval page	What it means
Signed in as	The Chat.co account the app will act for. If it is the wrong account, cancel, sign out of Chat.co and start again from the app.
Workspace and billing account	The one workspace this connection uses. Its experts, permissions and message credits apply. To use a different workspace later, disconnect and connect again.
This application will be able to	Ask your experts and receive answers, citations and supporting-document links. Search knowledge you are permitted to read and open the sources it returns. When you ask for one, an expert may also produce an assessment or other file, which the app receives as a download.
Cannot	Manage or delete your chatbots. The connector has no tools for editing knowledge, settings, team members or billing.
Connection details	The requesting app’s client ID and return address. Use them to confirm this is the app you started from.

Choose Approve connection to finish, or Cancel to stop. The only permission requested is chatco:read. It means the app cannot change your chatbots, knowledge, settings or account. It does not make the connection free of side effects: asking an expert saves the conversation in Chat.co and uses message credits, and a standalone search uses credits too, exactly as in the Chat.co app. Your questions, and the answers, excerpts and files you are permitted to receive, pass to the AI app and its provider.

5. Your first cited answer

Run these in a Codex session. HR Handbook is a made-up example; use the name of one of your own experts.

1
Discover your experts
prompt
```
Which Chat.co experts can I ask?
```
Expect: the experts in the workspace you approved, with short descriptions, and whether you can search each one's documents directly. Free.
2
Ask a question and request citations
prompt
```
Ask the HR Handbook expert how many vacation days a new employee gets. Cite the sources.
```
Expect: the expert's answer with citations that link to supporting documents. It is charged once, at the expert's normal Chat.co rate. Expert answers can take longer than an ordinary chat reply; while it waits, the app may check back on the same question several times, and you are still charged once. Answers do not include credit-cost notes; your usage appears in Chat.co.
3
Follow up in the same conversation
prompt
```
Does that change after two years of service?
```
Expect: an answer that builds on the previous one, because the app continues the same expert conversation. If the reply ignores the earlier context, ask the app to continue the same Chat.co conversation.
4
Open a source
prompt
```
Open the source for the first citation.
```
Expect: the cited passage in the conversation, or a link to a Chat.co supporting-document page. Private documents ask you to sign in with the account that received the answer. Opening sources is free, even when credits have run out.
5
Optional: search documents directly
prompt
```
Search the HR Handbook expert's documents for "parental leave".
```
Expect: a list of matching sources you can open. This works only for experts marked as directly searchable. Each completed search costs 0.5 message credit, even when nothing matches.

6. Message credits

Activity	Charge
Ask an expert (including follow-ups)	The expert’s normal Chat.co message rate
Completed standalone document search	0.5 message credit
Retrieval an expert performs inside its answer	No additional charge
List experts or open returned sources	Free

A search that completes with no matches still counts. Failed operations, inaccessible experts and retries of the same operation are not charged again. Your existing allocation applies; there is no checkout inside the connector. When credits run out, new paid requests stop, but you can still list experts and open sources you already received.

7. Common errors

What you see	Cause and fix
codex mcp add does not accept --url	Your Codex is older than remote-server support. Update Codex, then run the command again.
Tool calls fail with an authentication error	Run codex mcp login chatco again. The sign-in may have been revoked in Chat.co.
Codex answers without using Chat.co	Name it in the request: “Use my Chat.co experts to…”. Check /mcp shows chatco as enabled.
“This connection request expired or could not be verified.”	The sign-in link was stale or already used. Start the connection again from the app.
“No accessible workspace” on the approval page	This login has no workspace with expert access. Sign in with the right account, or ask the account owner for access.
An expert is missing from the list	You chose a different workspace, or your access or licence changed. Check the expert in Chat.co, then disconnect and reconnect with the right workspace.
“Expert unavailable.”	The expert is not available to this connection. Ask the app to list your experts again and pick one from that list.
“This expert has insufficient message credits…”	The allocation is used up. Listing experts and opening sources still work. Add credits through your usual Chat.co process.
“The selected account does not have an active subscription.”	The workspace’s Chat.co plan is not active. The account owner can check billing in Chat.co.
“The expert took too long. This operation was not charged.”	Try again, or split a very broad question into smaller ones.
Search returns nothing, or asks which expert should pay	Name one expert in the request. Direct search works only for experts marked as searchable.
“This document is unavailable or you no longer have access.”	Sign in to Chat.co with the account that received the answer. Removed access or a deleted source cannot be opened.

8. Disconnect

1
Revoke access in Chat.co
Open Connected applications in Chat.co, find the app and workspace, and choose Disconnect. The app loses access to the connector from its next request.
2
Remove the server from the app
Sign out and remove the server from Codex:
bash
```
codex mcp logout chatco
codex mcp remove chatco
```
To pause it instead, set enabled = false in its config.toml block. codex mcp --help lists the commands your version supports.

Disconnecting stops future access. It cannot remove answers or excerpts the app already received; that app's own terms and privacy policy govern them. To reconnect, start again from the app.

9. For developers & agents

You don't need this section to use Chat.co from an AI app. It documents the protocol for developers and for agents that call the tools directly.

Server

Property	Value
endpoint	`https://mcp.chat.co/mcp`
transport	MCP Streamable HTTP, stateless. POST only; GET and DELETE return 405 because there is no session to resume.
auth	OAuth authorization code with PKCE and dynamic client registration. Discovery starts from `https://mcp.chat.co/.well-known/oauth-protected-resource`, which the server also advertises in its `401` response.
scope	`chatco:read`
authorization	Re-checked on every call: current workspace membership, expert access, licence and credits.

Tools

Tool	Arguments	Returns	Cost
list_authorized_experts	None	`experts[]` with `id`, `name`, `description`, `access.chat` and `access.knowledgeSearch`	Free
consult_expert	`expert_id`, `question` (≤16,000 chars), `operation_id`* (16–128 letters, digits, `-` or `_`), `conversation_id`	`conversation_id`, `answer`, `citations[]` (`id`, `title`, `url`), and `files[]` (same fields) only when a requested expert action produced an assessment or file	Expert’s normal rate, once
search	`query`* (≤8,000 chars), `operation_id`*, `expert_id`, `paying_expert_id`	`results[]` (`id`, `title`, `url`)	0.5 credit per completed search
fetch	`id`*, `title`	`id`, `title`, `text` (bounded excerpt), `url`, `metadata` (`expert`, page range when known)	Free

* Required.

Recipe for agents

1
List experts first
Call list_authorized_experts. Only use id values it returns. Offer search only for experts whose access.knowledgeSearch is true.

Consult with a fresh operation_id

Generate a new identifier (a UUID works) for each new question. Keep the returned conversation_id.

consult_expert argumentsjson

{
  "expert_id": "<id from list_authorized_experts>",
  "question": "How many vacation days does a new employee get? Cite the sources.",
  "operation_id": "7f3b2c1e-5a9d-4e6f-8b0a-2d4c6e8f1a3b"
}

result (abridged)json

{
  "conversation_id": "<keep for follow-ups>",
  "answer": "New employees receive … [1]",
  "citations": [
    {
      "id": "<source id>",
      "title": "Employee handbook",
      "url": "<source link>"
    }
  ]
}

Follow up in the same conversation

Pass the conversation_id back with a new operation_id.

consult_expert argumentsjson

{
  "expert_id": "<same expert>",
  "conversation_id": "<conversation_id from the first answer>",
  "question": "Does that change after two years of service?",
  "operation_id": "c41e9d27-0b6a-4f38-9e15-7a2d8c3b6f90"
}

4
Read evidence with fetch
Pass the citation's id and its title. The title is shown in approval prompts; the id is the lookup key. The excerpt may not be the whole document.
fetch argumentsjson
```
{
  "id": "<source id from the citation>",
  "title": "Employee handbook"
}
```
5
Present the answer with its links
Show every cited source as a link to its url. Offer entries under files as downloads, not as citations. Don't add credit rates, model names or internal configuration to the answer.

Retries and identifiers

A paid call waits about 25 seconds for its result. If the expert is still working, the tool returns a “still running” error. Repeat the identical call with the same operation_id until the result arrives, without starting a new question. Do the same when your own request times out or the connection drops before a reply. A completed operation returns its stored result and is never charged twice.

If the server says the operation failed or expired (including “The expert took too long”), it was not charged; to try again, send the request with a new operation_id. An operation_id reused with different arguments is rejected as a conflict.

search without expert_id covers at most three experts and needs paying_expert_id; the results are not exhaustive. Prefer a targeted search.

Synthetic example: the HR Handbook expert, its answer and the angle-bracket values are placeholders. Use the identifiers and source links the server actually returns.

10. References

Chat.co experts in your AI apps — overview, status of every host, credits and privacy
Connected applications — review or disconnect
OpenAI: Model Context Protocol in Codex
Support: support@chat.co

Use your Chat.co experts in Codex

Key takeaways

1. Availability

2. Before you start

3. Add Chat.co to Codex

Add the server

Sign in

Check it

Or edit config.toml

4. Approve in Chat.co

5. Your first cited answer

Discover your experts

Ask a question and request citations

Follow up in the same conversation

Open a source

Optional: search documents directly

6. Message credits

7. Common errors

8. Disconnect

Revoke access in Chat.co

Remove the server from the app

9. For developers & agents

Server

Tools

Recipe for agents

List experts first

Consult with a fresh operation_id

Follow up in the same conversation

Read evidence with fetch

Present the answer with its links

10. References