Your files. Your agents. Your judgment.
Codex and Claude can produce more than you can comfortably keep in your head. Turn the useful parts into a connected project library: readable decisions, traceable feedback and a clear place to continue.
Public npm release: pending verification. These features are implemented in the local alpha. This page previews the standalone experience; it is not a download announcement.
From a demo to your daily work
Attach existing Markdown folders. Read linked tenets, decisions, research and designs; navigate folders, search documents, pin useful pages and export a document through browser printing.
Demo boundary: fixed shared samples keep private work off this public host.
Keep inline conversations, attributed replies, stable IDs and open-thread counts. Follow the reasoning through revisions and return to outstanding questions.
Demo boundary: private visitor feedback expires after 24 hours; it is a trial workspace.
Select comment IDs and export a bounded owner handoff. Your existing Codex or Claude session retrieves authorized feedback and replies through the local CLI or API.
Demo boundary: no live agent execution, model keys or paid inference.
Issue scoped, expiring reviewer or agent access; check and repair attachments; close or reopen threads; record acceptance of a source revision; back up and restore the local installation.
Demo boundary: visitors explore review, while owner operations and changing source files stay outside the public playground.
Supplied samples and temporary server-stored comments. No visitor project attachment or AI-provider keys. No sign-in required does not mean no server or edge metadata.
Documents and review history live in your installation. You control access and backups. Application access credentials are distinct from AI-provider API keys. The browser never writes source files.
Codex or Claude runs under your account and settings. Relevant content you give it may be processed by its provider. Local ReviewWithAI does not change the provider’s handling of your data.
A sha256 revision identity identifies the source bytes. Inspect the revision label and Download source in the demo. Matching a passage is not evidence that its meaning is unchanged; owner acceptance remains a separate decision.
Nine steps. One familiar working loop.
Keep Codex or Claude where you already use it. Add a reviewable project record alongside it. These diagrams are product illustrations; C-042 is an illustrative ID, not an executed-agent record.
Start with the verified npm package when it is publicly released. The current alpha is shared locally; there is no public install available yet. Supported alpha environment: Linux x86-64 with Node.js 24.19.0 or later within Node 24.
Register the folder where Codex or Claude already keeps your project documents. Your existing workflow stays familiar; ReviewWithAI gives you a browser view of its Markdown.
Continue in your usual Codex or Claude session. Ask it to save useful conclusions as structured Markdown instead of leaving every important answer buried in chat.
Start small: tenets for principles, design-decisions for choices and reasons, open-points for unresolved questions, and research or a low-level design for your key vertical. These are ordinary files you ask your agent to write, not documents generated automatically by this tool.
Ask your agent to add relative links between related documents and sections. Browse folders, search documents and follow those links to understand the evidence behind a decision.
Select a passage, ask a precise question and set a priority. Each conversation has a stable comment ID, so your feedback remains addressable when you return.
Choose the comments to address and set the scope. Give your existing agent session the exported owner handoff, or selected IDs with authorized CLI access. The browser never starts an agent for you.
Your agent edits the project files outside the browser and replies through its authorized interface. Inspect the change, check whether the comment still attaches to the right passage and explicitly repair it when needed. A repair does not mean you accepted the change.
Record acceptance of the reviewed revision, keep unresolved questions visible and give your next session a focused starting point. The record supports continuity; the quality still depends on your judgment and how you use it.
A small instruction with a clear scope
Example owner instruction, after granting the agent appropriate local access. C-042 is an illustrative ID; use the actual ID copied from your project.
Read comment C-042 in the research document. Check the source behind the claim, revise only that section, and reply with what changed and the evidence. Leave acceptance to me.
Forthcoming local release
The public npm package is pending verification. Send an email to request one release notification.
[email protected] · Opens a draft in your mail app; you decide whether to send it.
Make your good habits go further
The productivity opportunity comes from finding earlier reasoning, asking better questions and avoiding repeated context-setting. ReviewWithAI is a tool; its real power comes from how you use it.
Remote workspace connectivity remains planned. Automatic agent execution and a built-in model service are not part of this experience.
Try the review experience →