.svg)
DataDoe + OpenAI CodexOne table in ~/.codex/config.toml plugs any MCP server into the Codex CLI, the IDE extension and the ChatGPT desktop app. Point it at DataDoe and Codex builds on fresh Seller Central, Vendor Central and Amazon Ads data, and changes listings only once you switch writes on.
Codex keeps its MCP servers in ~/.codex/config.toml. Add one [mcp_servers.<name>] table with the server’s URL, plus a header if the server wants a key, then restart Codex. Four lines are the whole DataDoe entry; the second version keeps the key out of the file and adds two settings worth knowing. Pick a key to see its line.
/mcp and codex mcp list show, and how Codex labels each DataDoe call.url tells Codex this is a remote server over streamable HTTP, so nothing runs on your machine.datadoe-mcp-key header. Paste the key in the file, or name an environment variable with env_http_headers so the file holds no secret. Create and revoke keys in DataDoe under Integrations → MCP.prompt, Codex stops and asks before it runs any DataDoe tool. Useful while you are still deciding which Action types to switch on.[mcp_servers.datadoe]url = "https://mcp.datadoe.com/mcp/v1"http_headers = { "datadoe-mcp-key" = "YOUR_DATADOE_MCP_KEY" }startup_timeout_sec = 30
[mcp_servers.datadoe]url = "https://mcp.datadoe.com/mcp/v1"env_http_headers = { "datadoe-mcp-key" = "DATADOE_MCP_KEY" }startup_timeout_sec = 30tool_timeout_sec = 120default_tools_approval_mode = "prompt"
With the env version, run export DATADOE_MCP_KEY=… in the shell that starts Codex. A project’s own .codex/config.toml is merged too, but only in folders you have marked as trusted, so keep a personal key in the home-level file.
url for command and args, or run codex mcp add <name> -- <command>.--url, then run codex mcp login <name> to finish OAuth in the browser.The CLI, the IDE extension and the ChatGPT desktop app read the same config.toml, so you set DataDoe up once. Cloud tasks run on OpenAI’s machines and never see that file.
| Surface | Reads config.toml | How DataDoe gets in | Amazon data |
|---|---|---|---|
| Codex CLITerminal | Yes | Edit the file, or start with codex mcp add datadoe --url https://mcp.datadoe.com/mcp/v1 and add the header line by hand | Yes, in every session |
| Codex IDE extensionVS Code and editors built on it | Yes, the same file | Gear menu → MCP servers, or nothing at all if the CLI already has it | Yes, next to your open repo |
| ChatGPT desktop appCodex inside the app | Yes, the same file | Nothing extra once the file has the entry | Yes |
| Codex cloud tasksRun in OpenAI’s sandbox | No | OpenAI does not document MCP servers for cloud tasks yet | Not directly. Run data jobs locally, send code-only work to the cloud |
| ChatGPT on the webChat, not Codex | No | Through the DataDoe app instead, see DataDoe for ChatGPT | Yes, in chats |
Based on OpenAI’s Codex MCP documentation, October 2026.
Thirty seconds now saves you from a long Codex run that quietly fell back to guessing. The same three checks work for any MCP server you add.
codex mcp list
DataDoe should be on the list under the name you gave the table.
/mcp
Type it inside a running Codex session. It lists every connected server and the tools each one offers.
Show me my Amazon Seller Central sales for the last 7 days.
Real numbers mean you are connected. Codex will ask before the call if you set approvals to prompt.
The five problems we see most when people connect Codex to DataDoe, and the fix for each.
~/.codex/config.toml; a project file only counts in a trusted folder. Restart Codex after any edit.startup_timeout_sec. DataDoe’s docs use 30; the Codex default of 10 can be too short on a slow network.Codex is at its best on jobs that take an hour of reading, writing and testing. With DataDoe connected it works from your real schema and numbers instead of mocks. Three briefs to try, and what a run looks like.
Build a nightly job that pulls yesterday’s orders, settlements, ad spend and refunds from DataDoe, applies our unit costs from costs.csv and writes margin per SKU and marketplace to Postgres. Add tests and a README.
Plan. Five steps: schema, fee columns, cost join, job, tests.
DataDoe returns the settlement, order and ad table schemas.
DataDoe samples 30 days so Codex can see how fees are split.
Writes the job, the SQL and a cost loader for costs.csv.
Runs tests against the sample. 12 pass.
You review the diff and merge.
Codex used MCP to learn your data while it wrote the code. The finished job calls the DataDoe REST API on its own schedule, so it keeps running after the session ends.
Our repricer in pricing/ hard-codes FBA fees. Refactor it to use real fee estimates and competing offers from DataDoe, keep a 15% margin floor, and replay last month to show which prices would have changed.
Reads pricing/ and finds fees hard-coded in three places.
DataDoe returns fee estimates for the catalog.
DataDoe returns competing offers and Featured Offer prices.
Refactors fees into one module with a margin-floor check.
Replays last month: 61 price changes, 9 blocked by the floor.
You decide whether to ship it.
The refactor touches code only. When the new repricer later submits a price, that is a separate Action type you switch on yourself.
Rewrite the bullet points for the SKUs in drafts/bullets.csv. Validate every change with DataDoe as a dry run, show me the failures, fix what you can, and ask me before submitting anything.
Drafts new bullets for 40 SKUs from the CSV.
DataDoe dry run: 37 valid, 3 rejected for a missing attribute.
Fixes the 3 and runs the dry run again. All 40 valid.
Codex asks before the submit call. You approve.
DataDoe submits 40 updates and logs each payload.
Amazon applies listing changes within minutes to a few hours. Every one sits in the DataDoe Action log, with the payload that was sent.
Reading is on from the start. Anything that would change your Amazon account has to get past three separate gates, and the last one is Codex. You can add a dry run before it.
DataDoe is one MCP endpoint, so the key you made for Codex works anywhere that speaks MCP. These guides cover the setups closest to this one.
Open ~/.codex/config.toml and add a [mcp_servers.<name>] table. For a remote server, give it a url and any headers it needs; for a local one, give it a command and args. Restart Codex, then run codex mcp list to confirm it loaded. The CLI also has codex mcp add for the same job.
In the .codex folder in your home directory, so ~/.codex/config.toml on macOS and Linux. Codex also merges a .codex/config.toml inside a project, but only in folders you have marked as trusted. Keep personal keys in the home-level file.
Yes. The Codex CLI, the IDE extension and the ChatGPT desktop app share one config.toml, so a server added in one shows up in the others. In the IDE you can also add servers from the gear menu under MCP servers.
Not directly today. Cloud tasks run on OpenAI's machines and do not read your local config.toml, and OpenAI does not document MCP servers for them yet. Run jobs that need Amazon data from the CLI, the IDE extension or the desktop app.
Use env_http_headers instead of http_headers: map the datadoe-mcp-key header to the name of an environment variable, such as DATADOE_MCP_KEY, and export the key in the shell that starts Codex. You can revoke a key in DataDoe at any time under Integrations, MCP.
Only through DataDoe Actions, and only after you allow it. Every Action type is off until you switch it on, prices are a separate type, and the Amazon account has to be set to read and write. You can ask for a dry run first, Codex can ask before each tool call, and every Action is logged.
Yes. Seller Central, Vendor Central and Amazon Ads sit in one DataDoe schema with 100+ tables, so Codex can join a vendor ASIN's purchase orders with its ad spend in the same job.
Separately. DataDoe costs $97 a month after a 14-day free trial, with unlimited Seller, Vendor and Ads connections. Codex is billed by OpenAI through your ChatGPT plan or API usage.
We'll walk through your repo and the jobs you want Codex to own, on a demo account with real Amazon data.
Book a demoConnect Amazon to DataDoe, paste four lines into config.toml, and every Codex job starts from fresh Seller, Vendor and Ads data — with writes off until you say so.
We use cookies to improve your experience and analyze traffic. By clicking "Accept", you agree to our use of cookies. Read the Cookie Policy.