AIChatRender

Read your Claude Code or Codex CLI session in a VS Code panel, with the math nicely rendered. Scroll back, browse turns, while you compose your next response. Cmd+c in the rendered panel, then Cmd+v pastes the Markdown translation.

Unaffiliated. AIChatRender is an independent tool. It is not made, endorsed, or supported by Anthropic or OpenAI. It reads the session files those CLIs already write to your own disk; it uses no API and no account.

Requirements

• Claude Code or Codex CLI, run from this workspace. AIChatRender reads the session files they already write; there is nothing to configure in either.
• VS Code 1.85 or newer.
• Linux or macOS. Following the active terminal needs process introspection: /proc on Linux, pgrep/lsof/ps on macOS. On Windows, use WSL or a remote/container workspace — AIChatRender declares "extensionKind": ["workspace"], so it runs on the Linux side of a remote window. Run natively on Windows, the panel and manual transcript picking still work, but Follow Active Terminal Chat reports that it is unsupported rather than silently failing.

Commands

Cmd+F (Ctrl+F) with the panel focused opens VS Code's find widget over the rendered chat. It searches the text as displayed, which is why Search Chat Source exists: the panel draws \theta as θ, so only a search of the source can match \theta.

What it does

• Renders math with a vendored, pinned KaTeX: $$…$$ and \[…\] display, \(…\) and $…$ inline, with Pandoc-style guards so currency and $VAR don't trigger false positives.
• Renders a small Markdown subset — headings, lists, blockquotes, rules, emphasis, code, and tables with alignment and math inside cells. Raw HTML and links stay inert text, deliberately.
• Copies back out as Markdown source. Equations become their original LaTeX; tables keep their pipe syntax and alignment.
• Shows tool calls in the terminal's visual language: the tool and its argument, the first lines of its result, and a count of what was left unread. File edits report added and removed line counts and draw the diff in the editor's own diff colours.
• Shows the newest 100 user turns of assistant text and real prompts. Thinking blocks, subagent traffic, and harness bookkeeping are skipped.

Each of these is covered in full in docs/reference.md.

How it works

AIChatRender watches the session files the two CLIs already write to your disk: Claude transcripts under ~/.claude/projects/, Codex rollouts under ~/.codex/sessions/. It follows whichever was written most recently and switches automatically.

A session counts as this workspace's when its cwd is the workspace folder or any directory beneath it, so a CLI started in a subfolder is still followed. Follow Active Terminal Chat goes further and pins the panel to the transcript held open by your active terminal's process tree; when that terminal can't be matched — an idle CLI often closes its transcript — a picker lists the workspace's recent chats instead.

Nothing is installed into either agent. There are no hooks, no configuration changes, and no second copy of your history.

Security and privacy

This is an extension that reads your AI chat transcripts, so the scope is worth stating precisely.

• Read-only. It never writes under ~/.claude/ or ~/.codex/, and keeps no copy, inbox, or history of its own.
• No network, and no local server. No HTTP listener, socket, port, or CDN request. A loopback server would be reachable by other processes on the machine; there isn't one. KaTeX is vendored and served over VS Code's internal webview channel.
• No agent hooks. Neither CLI's configuration is modified.
• Nothing from a transcript is executed or made clickable. Raw HTML and links stay inert, tool output is inserted as plain text rather than parsed, and Codex wrapper programs are statically analyzed but never run.
• Codex discovery reads the first line of rollouts from your other projects to find the one belonging to this workspace. Only cwd and a timestamp are retained; nothing from a nonmatching rollout is logged or reaches the webview.
• Transcripts are sensitive files. A session can hold prompts, command output, diffs, paths, and any secret that ever appeared in them. AIChatRender keeps no extra copy, but it does put that content on screen — worth remembering when sharing a screen or recording.

These choices reduce the extension's attack surface. They do not sandbox the underlying agent or make the panel a tamper-proof audit log. The reasoning behind each is in docs/reference.md.

Known limitations

• Copying an equation currently brings extra text with it. The original LaTeX is there and correct, but it arrives wrapped in two other renderings of the same formula, so a pasted $$…$$ block needs the surrounding text trimmed by hand. Prose, lists, code, and tables copy cleanly; this affects math only.
• Exactly one workspace folder.
• With concurrent Claude and Codex sessions, the panel follows whichever matching file changed last unless you pin one.
• Codex discovery scans only the last seven date directories, so an older rollout is not followed even when resumed.
• These CLIs ship frequently and their session formats are internal and undocumented. A format change can break parsing until it's fixed; parsing is fail-soft, so an unfamiliar record leaves prose intact and raises the panel's format-drift notice.

The full list, including Codex code-mode and macOS PID-reuse caveats, is in docs/reference.md.

License and citation

MIT — see LICENSE. Use it freely, including in commercial work. Bundled KaTeX is separately MIT-licensed; see media/katex/LICENSE.

If AIChatRender is useful in academic work, a citation is appreciated but not required by the license:

Chen, Andrew Y. (2026). AIChatRender: rendered math for AI coding-agent transcripts in VS Code.