A desktop GUI for PR-Agent ยท A local, semi-automated AI code-review client for the individual reviewer
Graphical interface (GUI) for the community PR-Agent ยท Electron desktop app ยท All data stays on your machine
Keywords๏ผPR-Agent GUI ยท pr-agent desktop client ยท AI code review ยท Pull Request & Merge Request review ยท Bitbucket / GitHub reviewer tool ยท local / self-hosted / private deployment
English ยท ็ฎไฝไธญๆ ยท Website
Code Meeseeks (internal codename meebox) is a desktop graphical interface (GUI) for the command-line tool pr-agent: it packages AI-assisted code review into a desktop client โ it fetches the PRs (Pull Requests / Merge Requests) awaiting your review, runs pr-agent locally to generate review comments, and lets the reviewer confirm / edit each one before publishing it back to the code hosting platform (GitHub / Bitbucket / GitLab).
Core design stance:
- Human decides โ every comment must be confirmed / edited by the reviewer before it reaches the remote; the AI only drafts.
- Rules stay local โ the reviewer configures their own check rules, style preferences, and LLM provider.
- Data stays local โ repository mirrors, PR metadata, and comment drafts all live in a local working directory; friendly to corporate intranets.
Inspired by Mr. Meeseeks from Rick and Morty: summoned on demand, does exactly one thing, and vanishes once it's done.
- โ Engineers / Tech Leads who take on code-review duties
- โ Those who want AI to speed up review while keeping the final decision, rather than handing judgment entirely to a bot
- โ Teams running self-hosted Bitbucket / GitLab inside a corporate intranet
- โ Not a review bot that runs automatically in CI (that role belongs to pr-agent itself)
- โ Not a collaborative team review platform (no server, no multi-user sync)
- โ Not a replacement for the code platform's native review interface
- Unified access to GitHub / Bitbucket / GitLab โ for self-hosted GitHub Enterprise / GitLab Self-Managed, just enter the instance URL; capabilities adapt per platform (e.g. graceful degradation of GitLab CE/EE approvals).
- Local-first, works out of the box โ repository mirrors, PR metadata, and comment drafts all live in a local working directory, friendly to corporate intranets; pr-agent is embedded in the installer, with zero external dependencies.
- HTTP proxy โ LLM calls, code platforms, and git fetches all go through an HTTP proxy, with local addresses connecting directly.
- Automatic discovery โ polls for PRs awaiting review, categorized ("to review / created by me", etc.) and grouped by repository, with status filtering and search.
- Unread and mention markers โ new assignments / new commits / @-mentions / replies are marked unread, with mentions counted separately.
- History and open-on-demand โ browse merged / closed PRs, or open any PR directly by URL (including adding comments and re-running review).
- Side-by-side / inline diff โ an editor-grade reading experience: file tree (with merge-conflict markers), view by change range / single commit, scrollbar overview ruler, blame, and cross-file code search.
- Inline comments โ comment on both added and removed lines; selected code can be referenced as context in a question.
- Command-driven pr-agent โ drive it conversationally with
/describe,/review,/improve,/ask; results are structured into actionable review findings. - Re-review loop โ raise an
/askre-review on a review suggestion, and per the verdict (supersede / keep / withdraw) the original comment is automatically superseded or closed. - Agentic orchestration โ natural-language-driven planning + multi-tool orchestration + long-term Memory, with an observable process (a think โ tool โ think timeline); you can add input mid-run and stop at any time, turning review into collaboration with accumulating context.
- AutoPilot pre-review โ automatically pre-runs review on new PRs pending your review, so drafts are ready for confirmation the moment you open the app; write actions are gated by per-item authorization + red-line checks (read-only tools only by default).
- Confirm โ publish โ turn findings into drafts, edit inline, and publish to the remote one by one or in bulk; the remote mergeable state is visible, and you can merge with one click when conditions are met (or via
/merge). - Comment interaction โ reply to / edit / delete your own comments; supports emoji reactions, @-mention completion, image attachments (paste / pick to upload), and
:shortcode:emoji rendering (as each platform allows). - Activity timeline โ comments / commit updates / review verdicts are merged into a single timeline (GitHub / Bitbucket).
- Notifications โ categorized system notifications for new PRs, comment replies, and @-mentions, only for pending PRs and never for already-decided items; click to jump straight to the PR or code line, with a macOS dock badge showing how many comments await your response.
- Multiple LLM providers โ OpenAI / openai-compatible / DeepSeek / Anthropic / Tongyi Qianwen / Volcengine Ark, and more (local Ollama connects via the openai-compatible
/v1endpoint); you can also call third-party models through an authorized local CLI tool (e.g. claude / codex). - Personalized rules โ each reviewer maintains their own rules directory (markdown + frontmatter, organized recursively into subdirectories), matched by project / repository / target branch; matched rules are injected into the review in Ruleset sections, ordered by
priority. - Tunable runtime parameters โ review task concurrency, input context length, and agent strategy (auto follow-up toggle, number of follow-ups / code suggestions) are all adjustable on the settings page.
- Integrable by external agents โ PR review capabilities are exposed via a local HTTP interface + a cross-platform CLI, letting local agentic tools (e.g. claude / codex), scripts, and CI fold PR discovery / browsing / review-agent operations into automated workflows.
- Local API service โ optionally enable a local API that exposes PR discovery / browsing / diff / review-agent operations / review write actions over a language-agnostic HTTP contract; reachable only from localhost by default, with mandatory access-token auth, and no merge or mutating tools exposed.
- Cross-platform CLI
meeboxโ Windows / macOS / Linux clients ship with each release to browse PRs, drive the review agent, and perform review write actions (approve / needswork / comment) via the local API;meebox loginstores credentials once so you never pass them again, and the archive is itself an agent skill directory that drops straight into an agent's skills folder. See CLI tool for usage.
- Themes and appearance โ dark / light / follow system, several editor color themes, and a custom monospace font and size.
- Command palette โ invoke with
Ctrl/Cmd+Shift+Pto quickly run common actions and centralize scattered features. - Multilingual UI โ English / ็ฎไฝไธญๆ / ๆฅๆฌ่ช / Deutsch, with the AI's reply language following the UI language.
- Frameless window โ a custom-drawn title bar (VS Code style) showing the brand name and current PR title.
Download the installer for your platform from Releases:
| Platform | Artifact | Status |
|---|---|---|
| Windows x64 | code-meeseeks-<version>-win-x64.exe (NSIS installer) |
โ Available |
| macOS arm64 | code-meeseeks-<version>-mac-arm64.dmg |
โ Available |
pr-agent is embedded in the installer, so it works right after installation with no extra environment setup.
First launch on macOS: the installer is ad-hoc signed and not Apple-notarized, so Gatekeeper will block the app from an unknown developer. On first use, right-click the app and choose "Open", or go to "System Settings โ Privacy & Security" and click "Open Anyway"; after confirming once it launches normally.
Why it isn't notarized: this project is free and open-source software without a paid Apple Developer account (which notarization requires). The source is fully public, auditable, and buildable yourself; ad-hoc signing does not affect functionality or security.
- Configure the connection โ enter the code platform URL + credentials on the settings page.
- Configure the LLM โ pick a provider and fill in the API key / base_url / model name.
- Discover PRs โ the app automatically polls for PRs awaiting your review, grouped by repository in the left list.
- Read + review โ select a PR to view the diff, and click the auto-review button to have the AI generate findings; you can also enter fixed commands (e.g.
/describe) or a natural-language request in the dialog box. - Confirm + publish โ turn findings into drafts, edit the wording, and publish to the remote one by one or in bulk.
Configuration lives in ~/.code-meeseeks/config.yaml; repository mirrors default to ~/.code-meeseeks/repos/ and can be moved elsewhere on the settings page.
HTTP proxy (optional, for intranet users): under "Network proxy" on the settings page, enter the HTTP proxy host / port / Basic Auth; once enabled, LLM calls, code platforms, and git fetches all go through the proxy, with local addresses connecting directly (includes a "Test connection" button).
Git fetches over SSH do not use this setting; configure
ProxyCommandfor the relevant host in~/.ssh/configyourself.
For details on each step (installation and first use, PAT permissions and clone protocols, LLM model selection, HTTP proxy) see the User guide.
| Platform | Status |
|---|---|
| GitHub | โ Verified (github.com + GitHub Enterprise Server, REST API v3) |
| Bitbucket Server / Data Center | โ Supported (REST API v1, >= 7.0) |
| GitLab | โ Supported (gitlab.com + Self-Managed, CE / EE, REST API v4, >= 13.8, 15.6+ recommended) |
Review capabilities connect through pr-agent (backed by litellm), so it is in principle compatible with any OpenAI-compatible / litellm-supported model provider (pick the provider on the settings page and fill in the API key, base_url, and model name). The table below lists the built-in provider options on the settings page and their tested status:
| Provider | Notes | Status |
|---|---|---|
openai |
OpenAI (GPT family) | โ Verified |
anthropic |
Anthropic (Claude family) | โ Verified |
deepseek |
DeepSeek | โ Verified |
dashscope |
Alibaba Bailian (DashScope, Tongyi Qianwen) | โ Verified |
volcengine-ark |
Volcengine Ark (Doubao) | โ Verified |
openai-compatible |
OpenAI-protocol compatible (vLLM / gateway / self-hosted / local Ollama /v1) |
โ Verified |
cli |
Call a third-party model via a local CLI tool | โ Verified |
About local CLI mode: this mode does not call a model API directly; instead it hands the review request to a local command-line tool that you install and authorize yourself, which in turn calls the third-party model behind it. You must first install and log in to the corresponding CLI tool on your machine; the app itself does not manage its credentials or billing.
๐ธ Cost note: Agentic review and AutoPilot pre-review chain multiple model calls per PR (describe, review, follow-ups when needed, summarize), so token consumption is noticeably higher than a single manual review. Whether you use a pay-as-you-go API or a subscription / local CLI account with a quota cap, watch your usage pace and assess the cost yourself; per-step token usage is shown step by step on the review timeline for easy monitoring.
- Desktop shell: Electron + Vite (electron-vite)
- Renderer: React + TypeScript (strict)
- Editor: Monaco (side-by-side / inline diff)
- Engineering: npm workspaces + Nx monorepo
- pr-agent integration: an embedded Python-runtime subprocess (falls back to the system pr-agent CLI when absent)
๐ Further reading
- Delivered capabilities, roadmap, and open items: Roadmap
- Detailed architecture and per-module design: Module docs
For environment setup, running in dev, and building / packaging, see the Development guide.
- Local-first: apart from calling the LLM API and accessing the configured Git platform, no data is reported to any third party.
- Working directory: app data lives at
~/.code-meeseeks/, with a configurable repository mirror directory. - During review, pr-agent sends only the PR diff + the reviewer's rules to the reviewer's own configured LLM; wire up a local model (e.g. local Ollama) and nothing ever leaves your machine.
Built on top of PR-Agent โ the open-source version Qodo contributes to the community (site: docs.pr-agent.ai). It is bundled as a third-party dependency, distributed under its own license, and outside the renaming / modification scope of this project.
This project is licensed under the Apache License 2.0.
The distributed installers bundle third-party components (PR-Agent, Electron, etc.), each distributed under its own license, aggregated in NOTICE; the full third-party license set (THIRD-PARTY-NOTICES) is generated automatically at build time and bundled into the installer (in the app's resources directory).
This project is an unofficial, independent open-source tool, not affiliated with, authorized by, or endorsed by Rick and Morty or its rights holders in any way. Names, characters, and related elements such as "Rick and Morty" and "Mr. Meeseeks" are the copyrights and trademarks of their respective owners (Adult Swim / Warner Bros. Discovery, etc.). This project's name and icon are borrowed purely as an homage and assert no related rights; should the rights holders object, we will adjust accordingly.
Made on ๐ with