Plannotator: Local Browser Workspace for Reviewing AI Coding Agent Outputs and Feedback
Plannotator is an open-source local browser workspace designed to review plans, rendered HTML artifacts, and code diffs generated by AI coding agents such as Cl
As autonomous coding agents that write and edit source code directly inside terminal environments gain widespread adoption, a new operational challenge has emerged: how human developers can effectively review, annotate, and guide agent workflows without friction. Plannotator, an open-source project by backnotprop, provides a local browser-based review surface built to bridge this gap across agents such as Claude Code, Codex, Copilot CLI, Gemini CLI, and Antigravity.

Image source: https://github.com/backnotprop/plannotator
Reaching over 10,241 GitHub stars and releasing version 0.28.63 as of November 13, 2026, Plannotator focuses on eliminating the communication bottlenecks that occur when humans and autonomous coding tools collaborate.
The Bottleneck in Terminal-Based Agent Reviews
Terminal-based AI coding agents typically output extensive Markdown execution plans or large unified diffs directly into standard terminal output streams before touching code. For human reviewers, inspecting hundreds of lines of code modifications, formatting nuances, or interactive HTML prototypes within raw terminal scrollback is visually exhausting and error-prone.
When developers need to critique a specific line of code or an ambiguous section in an execution proposal, they often have to manually copy file paths and line numbers back into the prompt prompt box. This disjointed feedback loop frequently leads to context misunderstandings, causing the agent to misinterpret corrections and derail subsequent edits.
Plannotator solves this by hooking into agent runtimes. Whenever an agent formulates an execution plan or finishes drafting changes, Plannotator automatically opens a local browser workspace mirroring familiar code review interfaces. Reviewers can highlight text, write targeted inline notes, and return the aggregated comments directly into the agent's next message cycle.
Supported Ecosystem and Core Capabilities
Plannotator connects with AI harnesses via plugins, execution hooks, slash commands, and the Model Context Protocol (MCP). Supported tools include Claude Code, Codex, Copilot CLI, Gemini CLI, Antigravity, Kiro, Droid, Amp, and Pi, while its inbox layer accommodates assistants like Grok Bot, Muse, Hermes, and OpenClaw.
- Markdown Plan Review (Plan Mode): Hooks fire whenever an agent produces a plan. A clean Markdown review screen opens locally, enabling developers to drag-select passages, leave contextual notes, and dispatch feedback with a single click.
- HTML Artifact Rendering: Web prototypes, dashboards, and reports drafted by an agent render as live HTML rather than raw markup, allowing reviewers to visually inspect layouts and place feedback directly on top of components.
- Multi-VCS Code Diff Inspection: In addition to standard Git staging areas and branches, Plannotator handles remote GitHub pull requests, GitLab merge requests, GitButler workspaces, Jujutsu (
jj), and Perforce (p4) diffs. - Terminal Review Alternative (TUI): For workflows restricted to headless servers or developers who prefer not to leave the command line, Plannotator offers the Herdr Annotate plugin and a standalone Plannotator TUI. Annotations made in the terminal share the same underlying storage directory, keeping review history synchronized.
Command Workflow and Usage Examples
Once configured, interaction occurs through slash commands within supported agents or direct CLI invocations. In Codex environments, commands adapt to the !plannotator prefix or corresponding skills.
To inspect Markdown files, web URLs, or the previous response generated by the agent:
/plannotator-annotate README.md
/plannotator-annotate report.html --render-html
/plannotator-last
To review pending local workspace changes or remote pull requests:
/plannotator-review
/plannotator-review https://github.com/owner/repo/pull/123
plannotator review --gitbutler
To handle questions and tasks queued asynchronously across multiple agents:
plannotator inbox
The Plannotator Inbox collects pending decisions and questions in one local window without locking active terminal sessions. Once the reviewer enters replies and clicks send, feedback routes specifically to the individual agent session that initiated the query.
Privacy Posture and Implementation Considerations
Plannotator does not collect user analytics or telemetry. Plans, diffs, annotations, review history, and configurations remain on the local machine by default. However, engineering teams evaluating deployment should note three operational details:
First, whenever a review or archive surface loads, the app sends a standard HTTP query to GitHub to check for the latest Plannotator release. While no plan or review data is included in this query, there is currently no configuration flag to opt out of the update ping.
Second, the local Git review engine performs a git ls-remote call against the configured origin to check default branch alignment and flag stale diffs. The actual diff remains local. In air-gapped or restricted network setups, this remote query can be disabled entirely using plannotator review --no-git-remote-check, setting PLANNOTATOR_GIT_REMOTE_CHECK=0, or adding { "gitRemoteCheck": false } to ~/.plannotator/config.json.
Third, remote sharing architecture is transitioning. The legacy open-source asynchronous link-sharing feature is currently undergoing deprecation as team-based remote collaboration moves toward the hosted Workspaces offering. Teams operating under strict data boundary policies should rely on local single-machine workflows to maintain full control over internal code artifacts.
Sources
- GitHub Repository: backnotprop/plannotator