ChatGPT + Codex built 33 new DMR_REMOTE tools

Give ChatGPT hands. Then give it muscle.

I asked ChatGPT to find the agents working in my workshop. About 25 minutes later, it had a list: five main terminal agents, plus a detached Gemini session. The interesting part came next. FABLE already held useful agent information. That discovery became a specification, and Codex and ChatGPT turned it into a much stronger DMR_REMOTE MCP.

The result: 33 newly working tools in one session, with successful calls saved in the session evidence. Separately, the exposed catalog grew from 45 to 82 tools, an increase of more than 80%. One number counts successful new tool paths; the other counts available tool names.

Watch an AI team carry real work forward

The episode follows the actual workshop recordings: find the exact agent, send a question, receive a reply, offer work, receive a file and review the result. ChatGPT does more than describe the next step. It gets another agent involved and checks what comes back.

The tools that make DMR_REMOTE useful

• Read and search: fs_read_many gathers related files in one call; focused search finds the part that matters.
• Edit with checks: guarded edits can check that a file has not changed. Work Capsules hold prepared changes for review before publication.
• Run and inspect: run a task, track its progress, inspect the browser and look at images to check the result.
• Find and cooperate: exact agent identities, messages, work offers and result review connect ChatGPT with the agents already doing the work.

Jump to the moment

• 0:00 — ChatGPT’s agent hunt
• 2:00 — FABLE already had the information
• 3:04 — The DMR_REMOTE MCP blueprint
• 3:48 — 33 new working tools
• 5:01 — ChatGPT hands off real work
• 5:49 — The tools that give AI muscle
• 9:08 — DMR_REMOTE jumps from 45 to 82

The sequel to Give ChatGPT Hands

The original Chat On Steroids episode gave ChatGPT hands. This is the AI Powerlifter follow-up, recorded in my workshop on October 23.

Use DMR_REMOTE on your own computer

Update, October 24: a customer edition is now included with paid Workshop Access. It runs on Linux x64 or Windows 11 x64 through Ubuntu WSL2. Bring your own compatible AI account; ChatGPT web connections need your own HTTPS tunnel.

The episode shows my workshop setup. Its agent coordination needs a separate service that is not included in the customer bundle, and live desktop tools need a separate LCU installation. See what the customer edition includes before subscribing.

Limits of this report

• The 33 successful new tools and the 82-name catalog are different measurements; catalog exposure alone does not prove every operation.
• The recorded agent searches ran in different workshop states. Their timings are observations, not a controlled model speed ranking.