Codex MCP servers: setup and config

Codex MCP lets OpenAI Codex discover and call external tools through the Model Context Protocol. Codex can use local stdio servers or remote HTTP servers, including OAuth-protected endpoints. With AIPx, one agent URL exposes the connected account tools that you have permitted for that Codex workflow.

Add an MCP server to Codex

Connect the service accounts in AIPx first. Create an agent for Codex, decide which operations it needs, and copy its personal endpoint. Then register and authenticate the remote server:

codex mcp add aipx --url https://dash.aipx.run/mcp/AGENT_ID
codex mcp login aipx
codex mcp list

1. Replace AGENT_ID with the dashboard value.
2. Complete the browser login opened by the OAuth command.
3. Start Codex in the intended project and inspect /mcp.
4. Check that the tool list matches your agent’s permitted accounts.
5. Ask for one known GitHub issue or Sentry report before trying an operation that changes data.

The CLI registration uses a URL because AIPx is hosted. A local server registration instead describes the command that starts its process. Do not interchange those configurations or supply an upstream account key as if it were the AIPx MCP credential. OpenAI’s Codex MCP documentation covers remote transport and login.

For a first connection, keep the task narrow: retrieve the acceptance criteria for a named issue and report which account supplied them. That verifies the endpoint, authentication, account access and a useful tool call without creating an unintended write.

Codex MCP config: config.toml

Codex declares servers in TOML, rather than the mcpServers JSON used by several other clients. A minimal AIPx entry looks like this:

[mcp_servers.aipx]
url = "https://dash.aipx.run/mcp/AGENT_ID"

The user configuration is ~/.codex/config.toml. Trusted projects can also use .codex/config.toml. The CLI, IDE extension and supported desktop surfaces share the configuration model documented by OpenAI. Inspect the active project and configuration before diagnosing a server that appears in one session but not another.

A project file can communicate which endpoint the repository expects. Authentication is still personal. Never commit access tokens or upstream API keys beside the URL. The endpoint identifies an agent; the OAuth flow grants the client access to that endpoint. If teammates have separate AIPx accounts or task permissions, each may need a different agent declaration.

TOML section names and quoting matter. A mcpServers object pasted from Cursor is not valid Codex configuration. Keep the Codex-specific file and use Cursor’s setup guide or Claude Code’s configuration guide when configuring those clients. Sharing a server does not make their configuration syntax interchangeable.

OpenAI Codex MCP support

Codex supports external tool servers through stdio and Streamable HTTP, with OAuth for compatible remote servers. This is a capability of the client, separate from whether any particular service offers an MCP server. A REST API needs an adapter or a supported integration before Codex can use it as MCP tools.

Connection	What Codex needs	Typical fit
Local stdio	A runnable command and its arguments	Tools close to local files or a private environment
Remote HTTP	A reachable endpoint and supported authentication	Hosted app integrations
AIPx remote agent	The agent URL and OAuth login	Several accounts with different tool rules

A compatible transport is only the first check. The server must expose useful tool descriptions, input schemas and operations that your account can actually call. A client can connect successfully while particular tools remain hidden or the upstream provider denies access. Inspect each layer rather than assuming an MCP connection grants all service privileges.

Codex as an MCP client

When you add an external server to Codex, Codex is the client. It requests the tool list, chooses calls during the task and receives results. The upstream system remains responsible for executing its operations and enforcing its own account access.

codex mcp-server refers to a different direction: it exposes Codex to another MCP host. It is not the command for adding GitHub, Linear or AIPx to your Codex session. This distinction is useful when a search result says “Codex MCP server” but shows instructions for controlling Codex from a different application.

For ordinary coding work, leave the local repository operations with Codex’s existing tools. Connect an external server for the information outside the checkout: issue requirements, production errors, analytics evidence or service administration. After reading that context, Codex still needs to run the project’s checks before you trust a proposed fix.

Best MCP servers for Codex

Pick integrations according to the missing evidence in your development loop. Each link below covers its connection and tool boundary; the curated MCP server list supplies a wider comparison.

Integration	Useful input to a coding task	Access decision
GitHub	Issues, pull requests and repository context	Confirm mutations to the shared repository
Linear	A live issue brief and acceptance criteria	Start with reading assigned work
Sentry	Errors and their diagnostic context	Limit to the relevant project
PostHog	Evidence about feature behavior	Review analytics data exposure
Supabase	Database and project context	Use read-only project restrictions where possible
Cloudflare	Infrastructure context	Inspect available administrative tools
Jira	Requirements from a ticket	Hide unrelated projects or accounts
Confluence	Technical decisions and documentation	Keep document writes on Confirm
Notion	Product notes behind implementation work	Review workspace access
Google Sheets	Test matrices or rollout checklists	Confirm cell changes
Stripe	Billing integration context	Restrict financial write operations
Gmail	A requested work discussion	Hide personal mail from the coding agent

For a small repository, GitHub, Linear and Sentry may cover most external context. Adding calendar, health or payments tools without a relevant task increases what the agent can discover without improving the fix. A direct vendor server may be the best choice if you need only that vendor and prefer its native setup.

Give Codex scoped access

In AIPx, rules combine the agent, connected account and tool. A code-maintenance agent can read the work tracker, ask for confirmation before opening an issue, and have no access to a personal account. Another agent can use the same connected account with different rules.

Allow and Confirm are execution hints, with Confirm requesting client-side review. Actual confirmation depends on the client; Hide is enforced server-side and withholds the operation for that agent. Review reads as well as writes. A read of customer records or private mail can disclose data to the conversation even though it changes nothing upstream. AIPx’s defaults cannot know your confidentiality requirements.

The audit records account, agent, tool, parameters, result and time. Review the first task there, especially if several accounts have similar names. Codex’s own approval and sandbox controls remain relevant to local execution; gateway permission rules govern the external account tools. The MCP gateway guide explains how those accounts are shared across clients.

Troubleshooting Codex MCP

If a server is absent, check the TOML table, project trust and active configuration. If login fails, repeat codex mcp login aipx and verify that the browser signs into the intended AIPx account. If the connection works but a tool is absent, review the agent’s Hide rules and refresh discovery. If the provider rejects a call, reconnect or narrow the upstream account access instead of guessing at a client token.

AIPx adds another hosted operator and depends on network and provider availability. It is not a deployment target for your own local process. Prefer a local server when the tool must remain in a private environment.

Connect your tools to Codex

Beyond the development shortlist, these supported services can supply infrastructure, meeting or specialist context. Keep personal accounts separate from the coding agent unless the task explicitly needs them.

• Development and infrastructure: Atlassian Rovo tools, DigitalOcean infrastructure.
• Mail, notes and planning: Google Calendar events, Apple Calendar access, Things 3 tasks, Granola meeting notes, Fireflies transcripts, AgentMail inbox tools.
• Payments: Paddle subscription tools.
• Research and campaign data: Exa web search, Google Maps data, X account tools, Yandex Direct campaigns.
• Personal health and nutrition: WHOOP recovery data, Withings measurements, Hevy workouts, Numon nutrition reports.

Free includes three connected accounts and one agent. Pro is $3 per month with unlimited connected accounts and agents plus Pro integrations; upstream subscriptions and API limits still apply. Connect free.