pi-gpt

A Pi extension that lets your agent use your ChatGPT account — start chats, run deep research, analyze code, and continue conversations — all from inside Pi. No API key needed: it reuses your codex login.

Install

pi install -l .               # from the pi-gpt directory
# or from anywhere:
pi install -l /path/to/pi-gpt
# or, once published:
pi install npm:pi-gpt

Requires a ChatGPT login:

codex login   # writes ~/.codex/auth.json

What you get

| Tool | What it does | |------|-------------| | gpt_account_status | Shows your plan, quota (limit, used, reset time in your timezone), and available workspaces. | | gpt_list_models | Lists available ChatGPT models and their thinking-effort options. | | gpt_chat | Starts or continues a conversation. Supports normal chat, agent mode, and deep research. Accepts files and prompt files. | | gpt_list_chats | Lists chats started from the current project. | | gpt_get_conversation | Reads the full message history of a conversation. | | gpt_get_message | Reads a single message by ID. |

Chat types

| Type | When to use | Typical time | |------|------------|-------------| | normal | Quick questions, code review with files, general chat | seconds | | agent | Multi-step tasks, tool use, deeper analysis with files | seconds–minutes | | deep_research | Web research with citations (lighter model) | minutes | | deep_research_heavy | In-depth research with GPT-5.5 Pro + web browsing | 5–30 min |

Intelligence levels

For normal chats: instant → medium → high → extra_high → pro. Higher levels use more capable models and deeper thinking. pro uses gpt-5-5-pro. You can also pin a specific model and thinking_effort (min / standard / extended / max).

Files

Pass local files as context — the agent can't read your filesystem directly:

files: ["src/auth.ts", "tests/auth.test.ts"]
prompt: "Review these files for bugs."

• Text/code/Markdown: inlined into the prompt (all chat types).
• Images (PNG/JPEG/GIF/BMP/WebP): uploaded (normal/agent only).
• PDF: uploaded (normal/agent only).
• Deep Research: text only — images and PDFs are rejected before starting.

You can also use prompt_file to load a prepared prompt from a file.

Deep Research

Heavy deep research (deep_research_heavy) runs a multi-minute web research session using GPT-5.5 Pro. The tool handles everything automatically:

• Auto-starts: the research plan is confirmed immediately, no manual approval.
• Live progress: streams progress updates so you can see it working.
• Sources: citations with URLs are captured automatically.
• Timeouts: defaults to a 2-hour wait. On timeout, you get the partial result plus the conversation_id to retrieve the full report later.

Deep Research quota

Your ChatGPT plan includes a limited number of deep research requests per ~24-hour window. Both light (deep_research) and heavy (deep_research_heavy) share the same quota counter:

| Plan | Limit per window | |------|-----------------| | Pro | 40 |

When you run out, the tool tells you:

• How many you've used and the plan limit
• Exact reset time in your local timezone
• How many minutes until reset

Check your current usage anytime:

gpt_account_status
→ deep_research: { remaining: 12, limit: 40, reset_after_local: "Jul 28, 2026 at 2:25 AM" }

If a deep research request is rate-limited, the tool returns immediately with the reset details in details.deep_research_rate_limit — it does not wait or poll until the quota resets. Don't retry until the reported time.

Continuing conversations

gpt_chat returns a conversation_id. Pass it back to continue the same thread:

# First message
gpt_chat(prompt: "Explain JWT", chat_type: "normal")
→ conversation_id: "abc-123"

# Follow-up
gpt_chat(prompt: "How does refresh work?", conversation_id: "abc-123")

Multiple accounts

Give each ChatGPT login its own directory:

CODEX_HOME="$HOME/.codex-pro" codex login
CODEX_HOME="$HOME/.codex-work" codex login

CODEX_HOME="$HOME/.codex-pro" pi   # launch Pi with the Pro account

pi-gpt never switches accounts automatically. Call gpt_account_status to see which account is active.

Privacy

• Tool output is PII-redacted (emails, phone numbers, API keys, tokens).
• Files are contained to your working directory (path traversal blocked).
• No credentials are logged.

License

MIT.