Here is the honest version. The ChatGPT chat window cannot reach the files on your computer, so it cannot edit your vault on its own. What it can do is draft the exact commands, and OpenAI's Codex agent can read and write your vault directly. Your tasks are plain markdown, so either path files them in your format with no plugin and no API of your own to build.
Coming in TaskForge 3.0
The TaskForge CLI and MCP server ship with the TaskForge 3.0 release. Everything on this page describes how they will work at launch. The TaskForge app is available today.
Codex, OpenAI's coding agent, works like Claude Code: point it at your vault and it edits the files. If you would rather stay in the ChatGPT chat window, let it write taskforge commands and run them yourself. Both end with a real task in your vault.
Open your vault in TaskForge 3.0 or later, then Settings, AI & Agents, and keep "Generate agent files" on. TaskForge writes a taskforge/ folder describing your task format, which both paths below use so tasks land in your dialect.
Codex runs in your terminal and edits files in its working directory. Open your vault there, tell it to read taskforge/agents.md, and it manages tasks the way Claude Code does, matching your exact format.
cd ~/Documents/MyVault codex "Read taskforge/agents.md, then add the follow-ups from today's note to tasks/this-week.md, due Friday."
Install the taskforge CLI, then ask ChatGPT to write the command for what you want. It knows the taskforge query grammar well enough to draft it; you paste it into your terminal and the edit is real.
brew install azhard/tap/taskforge # then run what ChatGPT drafts, for example: taskforge add "call the plumber" --due 2026-07-10 --priority high
Either way, the task is now a markdown line in your vault. Open TaskForge and it is there, on your phone too once your vault sync runs.
The taskforge/ folder is what keeps a task in your format whether Codex writes it or the CLI does. Codex reads agents.md like any source file; the CLI reads schema.json to know your dialect. Neither guesses.
| taskforge/agents.md | Your task syntax in plain English, the file Codex reads before it writes. |
| taskforge/schema.json | The machine-readable version the taskforge CLI reads to match your format. |
| taskforge/lists.json | Your saved lists and where new tasks land, so a task filed to a named list ends up in the right note. |
Codex edits the file for you; with the CLI, ChatGPT drafts the command and you run it. Same result: a task in your vault, in your format.
You ask
“Codex: read meeting-notes.md and add every action item to tasks/this-week.md, due Friday.”
What lands in your vault
- [ ] send the signed contract to legal 📅 2026-07-10
- [ ] book the venue deposit 📅 2026-07-10
Codex reads the note and the contract, then writes both lines in your vault's syntax and shows the diff.
You ask
“ChatGPT: write me a taskforge command to list every urgent task, most urgent first.”
What lands in your vault
taskforge query 'prop:priority_label=urgent' --sort priority
ChatGPT drafts the command from the query grammar; you run it and get the list. Nothing about your vault ever leaves your machine to produce it.
Codex and the taskforge CLI both run on your machine and touch only the vault you point them at. The one thing that reaches OpenAI is the text of your conversation, the same as any ChatGPT chat. There is no TaskForge account and no cloud copy of your tasks.
The ChatGPT chat window cannot reach the files on your computer, so no, not on its own. OpenAI's Codex agent can, because it runs locally and edits files. Or you let ChatGPT draft taskforge CLI commands and run them yourself. Both end with a real edit in your vault.
The taskforge MCP server runs locally over stdin and stdout, which is the setup Claude Desktop and Cursor use. ChatGPT's connectors are built for hosted services, so the reliable OpenAI paths for a local vault are Codex and the CLI. We would rather tell you that than sell you a connector that does not fit.
Codex is OpenAI's coding agent. It runs in your terminal, reads and writes files in a working directory, and follows instructions like taskforge/agents.md. For managing vault tasks it behaves much like Claude Code.
Yes. Whether Codex or the CLI writes the line, it is a plain markdown task in your vault. TaskForge reads it on that device right away, and it reaches your phone after your vault syncs.
Both read the contract TaskForge writes into your vault. Codex reads taskforge/agents.md; the CLI reads taskforge/schema.json. That is why the task comes out in your dialect instead of a generic checkbox line.
The CLI does the heavy lifting for the ChatGPT path, and the AI overview covers every agent. If you use Claude or Cursor, those have dedicated walkthroughs too.
Plain markdown, edited by the tools you already use. TaskForge is free to download on iPhone, iPad, Mac, and Android.
Get Free - iOS & Mac 
Get Free - Android