We open sourced PlanBridge, a CLI tool for precision feedback on your coding agent's plans. It uses standard coding agent hooks (or skills) to open a local browser with the rendered markdown plan, letting you select text and leave inline comments on the plan.
Like most engineers, we use coding agents daily, but we realized the UX sucks for precise feedback on plans and other markdown. Reviewing even a short markdown plan in a terminal is tedious and frustrating. You have to scroll up and down, keep context in your head, and dump all your feedback into a single chat input at the bottom.
Because the friction is so high, we found ourselves settling for "good enough" plans. But every shortcoming or missing context in the plan becomes harder and more expensive to fix after the code is generated. Basically, we were burning time and tokens trying to clean up the code slop that was a direct result of imprecise plans.
So we built PlanBridge to make it frictionless to give precise feedback to a coding agent plan (and any other markdown) with document-style comments in a local browser.
Here is how it works under the hood:
1. It runs entirely locally. Your plan content never goes to a remote backend.
2. It uses standard harness hooks to get notified when an agent creates a plan (in "plan mode"). For Claude Code, it uses the plugin hook on ExitPlanMode. For Codex, it uses the Stop hook entry.
3. It pops open a local browser window with the rendered plan making it easy to read. You can then highlight specific text, drop inline comments exactly where you need to, and request changes (or approve if you have no feedback).
4. It pipes your structured, anchored feedback back to the agent, which then revises the plan. If you approved the plan, the agent starts coding.
5. For reviewing content that isn’t from a hook or outside of plan mode, you can use the /planbridge:open (for Codex its $planbridge-open) skill to iterate and give precise feedback.
We’re developers and, like you, hate invasive tools. PlanBridge is designed to fit into your existing coding agent workflow (no heavy UX) and uninstalls easily if you don’t find it useful.
Right now, the auto-installation supports Claude Code and Codex. If you use other agents, you can still use it asking your agent to run contextbridge for you.
You can try it out immediately (macOS/Linux): /bin/sh -c "$(curl -fsSL https://downloads.contextbridge.ai/cli/install.sh)"
Github: https://github.com/contextbridge/planbridge
Official Site: https://plan.contextbridge.ai/
Slack Community: https://go.contextbridge.ai/join-community/
We think feedback is a gift, so would love to hear yours. We’ve got some interesting features and ideas in the works including: support for precision code review feedback, allowing other agents to add comments to markdown files, summarizing what changed between plans, and lots of others in the issue tracker.
No comments yet.
