ipyai

ipyai is a terminal IPython extension with four AI backends:

• Codex API (codex-api, default) — hits the Codex responses endpoint directly using your ~/.codex/auth.json token
• Codex (codex) — local Codex app-server
• Claude CLI (claude-cli) — drives the claude -p CLI, so usage counts against your Claude subscription rather than API billing
• Claude API (claude-api)

It is aimed at terminal IPython, not notebook frontends.

Install

pip install -e ipyai

ipyai uses safepyrun for live Python state. Backend requirements:

• codex-api: local codex login so ~/.codex/auth.json holds a valid access token
• codex: local Codex app-server access
• claude-cli: local claude CLI install and Claude Code login (subscription auth)
• claude-api: Anthropic API access

How to Use Prompts

There are several ways to send an AI prompt from ipyai:

Dot prefix (.) — In normal IPython mode, start any line with . to send it as a prompt. Everything after the dot is sent to the selected backend. Continuation lines (without a dot) are included too, so you can write multi-line prompts:

.explain what this dataframe transform is doing

.draft a plan for this notebook:
focus on state management
and failure cases

Prompt mode — When prompt mode is on, every line you type is sent as an AI prompt by default. To run normal Python code instead, prefix the line with ;. Shell commands (!) and magics (%) still work as usual. There are three ways to enable prompt mode:

• opt-p (Alt-p) — toggle prompt mode on/off at any time from the terminal
• -p flag — start ipyai in prompt mode: ipyai -p
• prompt_mode config — set "prompt_mode": true in config.json to always start in prompt mode

You can also toggle prompt mode during a session with %ipyai prompt.

CLI

ipyai

Flags:

ipyai -r                 # resume last session for the selected backend
ipyai -r 43              # resume session 43
ipyai -l session.ipynb   # load a saved notebook session at startup
ipyai -b claude-cli      # select backend: codex-api | codex | claude-cli | claude-api
ipyai -p                 # start in prompt mode

On exit, ipyai prints the session ID so you can resume later.

Load As Extension

%load_ext ipyai

Usage

ipyai is a normal IPython session — you can run Python code exactly as you would in plain IPython. On top of that, you can send prompts to the selected AI backend as described above. %ipyai / %%ipyai magics are also available.

Useful commands:

%ipyai
%ipyai model sonnet
%ipyai completion_model haiku
%ipyai think m
%ipyai code_theme monokai
%ipyai log_exact true
%ipyai prompt
%ipyai save mysession
%ipyai load mysession
%ipyai sessions
%ipyai reset

Context Model

For each AI prompt, ipyai sends:

• recent IPython code as <code>
• string-literal note cells as <note>
• recent outputs as <output>
• the current request as <user-request>
• referenced live variables as <variable>
• referenced shell command output as <shell>

Prompt history is stored in SQLite; for compatibility, the table is currently named claude_prompts. Session metadata is stored in IPython's sessions.remark JSON, including cwd, backend, and provider_session_id.

Tools

ipyai exposes the same custom tools across all backends:

• pyrun: run Python in the live IPython namespace
• bash: run an allowed shell command via safecmd
• start_bgterm: start a persistent shell session
• write_stdin: send input to a persistent shell session and read output
• close_bgterm: close a persistent shell session
• lnhashview_file: view hash-addressed file lines for verified edits
• exhash_file: apply verified hash-addressed edits to a file

When using the Claude CLI backend, it also enables these built-in Claude Code tools:

• Bash
• Edit
• Read
• Skill
• WebFetch
• WebSearch
• Write

Custom tools are exposed to claude -p through an in-process unix-socket MCP server plus a small stdio bridge subprocess (ipyai-mcp-bridge), so the subscription-driven CLI can still call live-kernel tools like pyrun.

The ipyai CLI loads safepyrun before ipyai, so pyrun is available by default in normal terminal use. bash, start_bgterm, write_stdin, close_bgterm, lnhashview_file, and exhash_file are seeded into the user namespace by ipyai.

Skills

When using the Claude CLI backend, ipyai enables the built-in Skill tool and passes --setting-sources user,project so claude -p loads your normal user- and project-level skills.

Notebook Save/Load

%ipyai save <filename> writes a notebook snapshot. It stores:

• code cells
• note cells
• AI responses as markdown cells
• prompt metadata including both prompt and full_prompt

%ipyai load <filename> restores that notebook into a fresh session.

ipyai -l <filename> does the same during startup.

Backend restore is backend-specific:

• claude-cli: synthesizes a fresh Claude transcript JSONL each turn and claude -p --resumes from it
• claude-api: reuses the saved local prompt history directly on each turn
• codex-api: reuses the saved local prompt history directly on each turn (same flat-history flow as claude-api)
• codex: starts a fresh thread and sends the loaded notebook as XML context once

Keyboard Shortcuts

• Alt-.: AI inline completion
• Alt-p: toggle prompt mode
• Alt-Up/Down: history navigation
• Alt-Shift-W: paste all Python code blocks from the last response
• Alt-Shift-1 through Alt-Shift-9: paste the Nth Python code block
• Alt-Shift-Up/Down: cycle through extracted Python blocks

Config

Config lives under XDG_CONFIG_HOME/ipyai/:

• config.json
• sysp.txt
• exact-log.jsonl

config.json supports:

{
  "backend": "codex-api",
  "models": {
    "claude-cli":  {"model": "sonnet",                   "completion_model": "haiku",                   "think": "m"},
    "claude-api":  {"model": "claude-sonnet-4-6",        "completion_model": "claude-haiku-4-5-20251001","think": "m"},
    "codex":       {"model": "gpt-5.4",                  "completion_model": "gpt-5.4-mini",            "think": "m"},
    "codex-api":   {"model": "gpt-5.4",                  "completion_model": "gpt-5.4-mini",            "think": "m"}
  },
  "code_theme": "monokai",
  "log_exact": false,
  "prompt_mode": false
}

Development

See DEV.md.