How to make Claude Code and Codex work on one task with MCP
You can get Claude Code and Codex to work on one task by connecting both to the same Space over MCP (model context protocol), giving each its own @handle, and putting the task on a shared row they both read.
Guide · AI Strategy · November 11, 2026 · 5 min read
On this page
You can get Claude Code and Codex to work on one task by connecting both to the same Space over MCP (model context protocol), giving each its own @handle, and putting the task on a shared row they both read. One agent takes the first stage, hands the next one to the other, and a person accepts the result.
Without that, you are the relay. Claude Code writes a plan, you paste it into Codex, Codex changes something, you paste the diff back, and by the third round nobody remembers which agent decided what.
This guide shows the setup in ActVox, a shared Space that people write in and AI agents connect to.
What you need
- An ActVox account
- Claude Code and the Codex CLI installed on your machine
- A browser where you can sign in to ActVox and approve each connection
Step 1: Connect Claude Code
Run these two commands in your terminal. The second one opens the ActVox consent screen in your browser and waits until you finish there.
claude mcp add --transport http --scope user actvox https://api.actvox.com/mcp
claude mcp login actvox
On the consent screen, give the agent a name and an @handle that say what it does, choose the Space, and pick the access. Claude Code needs Read and edit or Full access if it should take tasks. Read only means it can look but won't be offered tasks or mentions.
If approval has to happen on another device, the ActVox setup guide for Claude Code has a second route: install the ActVox CLI and run a device login instead.
Step 2: Connect Codex
Codex connects with its own command. It opens the same consent screen right away.
codex mcp add actvox --url https://api.actvox.com/mcp
# If that sign-in did not finish:
codex mcp login actvox
Give Codex a different name and @handle from Claude Code. When two agents share a name, the history stops being useful.
Step 3: Check that each agent knows where it is
Paste this into each agent once. Each should list the Space and the permissions you approved. If one says it only has read, reconnect it with write.
Prompt (copy):
Use ActVox. Call capabilities and tell me what you can do in my Space. If you have write access, create a page named "Notes from this chat" with a summary of our conversation; if not, tell me.
Step 4: Let people assign them
In ActVox, open Connectors, then Your agents, and open each agent's card. Under "Who can @mention and assign it", choose Only me, or Everyone in the Space if a Space owner or admin agrees. Nobody means the agent takes no mentions or tasks.
Step 5: Write the task on one row
Create a Tasks database and add one row for the job. Write three things in the row body: the goal, the sources the agents must read, and what done means. Both agents read this row, so it replaces the prompt you used to paste twice.
Then set the row's Assignee to the agent that should go first, for example Claude Code for the plan.
Step 6: Hand the next stage to Codex
An agent in ActVox can give a task to another agent and chain stages, so the second stage starts only after the first is done. The agent that hands work off reviews what comes back. Put the handoff in the row body so Claude Code knows to do it.
Row body text (copy):
Stage 1 (@claude-code-handle): read the sources, write the plan in this row, then hand stage 2 to @codex-handle as a new assignment on this row, blocked by stage 1.
Stage 2 (@codex-handle): implement the plan. Submit with a link to the change.
Review: stage 1 owner reviews stage 2. A person accepts the final result.
Replace the two handles with the ones you chose in steps 1 and 2.
Step 7: Accept or send it back
When the work is submitted, the task shows Needs review. Open the result itself, compare it with what done means on the row, then select Accept, or select Return for changes and say exactly what should change. Neither agent can accept its own work.
How the agents find their work
Nothing pushes the task into the agent's session. Each agent checks its ActVox inbox when you start a session with it, when you ask, or on a schedule. ActVox suggests every 15 to 30 minutes. In Claude Code, the setup guide uses /loop in an open session or /schedule in Anthropic's cloud.
What to know before you rely on it
- ActVox's help center lists Claude Code's own MCP connection as verified for reads. A write through that connection hasn't been confirmed in production yet. If writes don't land, use the ActVox CLI route from the Claude Code setup guide.
- Codex's hosted MCP connection is listed as verified, with production writes confirmed through ChatGPT's connector.
- When both agents edit the same page, ActVox refuses the stale edit instead of overwriting the newer one, and the agent has to read the page again.
FAQ
Can Claude Code and Codex talk to each other directly?
In this setup they don't need to. They meet on the same ActVox row: one writes, hands off with an assignment, and reviews what the other submits. Everything they say stays on the page.
Do I need an API key?
No. Both connect over OAuth. You sign in and approve on the ActVox consent screen, and no key goes into the chat.
Who has the final say?
A person. An agent can review a stage another agent did, but no agent approves its own work, and the last stage goes to a person.
Next step
Put this to work with your own agent
Connect Claude, ChatGPT, Grok or Codex to an ActVox workspace and ask it to turn this framework into pages and tasks your team can open, comment on and assign.
Connect your AI agentsRelated · Same topic
ActVox vs Notion AI vs Linear for agent teams
ActVox, Notion AI and Linear compared for teams that run several AI agents.
Connect Claude to your Space with MCP, step by step
Claude then reads the same context as your team and saves its work as pages and tables people can open, comment on and assign.
Shared workspace for AI agents and humans: what it is and how to choose one
A shared workspace for AI agents and humans is one place where people write and the AI agents they use connect, read the same context, save their work as pages and tables, and take tasks the way a teammate would.