The agent says it fixed the checkout bug. You still need to find the pull request, read the diff, check the build, and open the right preview. Tomorrow, you need to remember which conversation contains the fix.
That last part is easy to underestimate. A useful agent session is a record of decisions, failed attempts, commands, and follow-up questions. Losing it means reconstructing work you've already done.
I built Cogpit to keep that work together. It's a free, open-source GUI for Claude Code, OpenAI Codex, and GitHub Copilot CLI. It runs on your machine and uses your existing CLI logins.
Here's the workflow I want it to make easier, using a checkout bug as an example. The screenshots show the current app with demo sessions and sample GitHub, ClickUp, and Vercel data.
Start with the task, keep the conversation
Say rapid clicks on "Add to cart" sometimes lose an item. The ClickUp task has the reproduction steps and the cases that need testing.
Cogpit's ClickUp panel lets you browse tasks and add a task to the prompt. The requirements stay beside the conversation while the agent works. You can open a command's output or a file edit without scrolling through pages of terminal text.
This also matters when you come back later. Cogpit can discover and resume existing Claude Code, Codex, and Copilot CLI sessions, including work started outside the app. Sessions are grouped by project, and you can search for a prompt, branch, or pull request.
You can keep sessions from all three CLIs in one workspace. Each conversation still belongs to its selected agent. I don't want switching projects to mean starting the explanation again.
Read the changes before trusting the summary
"Tests pass" is useful. It doesn't tell you whether the test covers the bug you reported.
For the cart example, I'd look for an assertion about the final cart contents after responses arrive out of order. A longer timeout might make a test green while leaving the race intact.
Cogpit lets you inspect edits by turn or review the session's combined diff. Its GitHub panel brings pull requests, changed files, reviews, and Actions runs into the same workspace. A linked session gives you a route back to the conversation behind a PR.
The useful question is specific. "Show me the regression test for a late cart response" gives the agent something concrete to answer, and gives you something concrete to review.
Open the page with the agent
A green build still leaves the browser check.
The Vercel panel shows production and preview deployments, with build logs you can expand. Find the preview for the branch you're reviewing, then inspect the page in the Browser panel.
The Browser panel streams a selected tab from an agent-browser browser. You can watch the agent navigate, click and type in the page yourself, or take over a login before handing it back.
The shared default browser keeps its logins between runs. Named browser sessions have separate persistent profiles, so a project can keep its own browsing context. You and the agent can select different tabs, so check that you're looking at the same one when discussing a page.
For this example, the check is simple. Add items quickly. Change quantities. Reload. See whether the cart still contains what you put in it. If it fails, the conversation is right there.
Try it on a session you already have
Install and sign in to Claude Code, Codex, or GitHub Copilot CLI, then download Cogpit for macOS, Windows, or Linux. Windows support has had less testing; the Browser host currently runs on macOS and Linux.
Or open the local web app with Node.js 20.11 or newer:
npx cogpit@latest
Pick an existing session and inspect one file it changed. That's enough to see whether this way of working fits you. GitHub, ClickUp, Vercel, and browser automation need their own setup; the documentation covers it.
If you try it, I'd like to know where you still lose context while working with coding agents. An issue with a concrete example is more useful to me than a feature wish list.
Top comments (0)