Run Claude Code on GitHub issues and pull requests

Install ProPR on a Linux Docker host, sign in with your Claude account during setup, and add a model label such as llm-claude-opus55 to a GitHub issue. Claude Code implements the issue in an isolated workspace, ProPR opens the pull request, and every later comment on that PR becomes a Claude Code follow-up commit.

Claude Code on GitHub issues: the setup

ProPR runs the official Claude Code CLI (@anthropic-ai/claude-code) on your Claude subscription or an Anthropic API key. You pay Anthropic directly; ProPR never marks up tokens. Claude Code in ProPR.

npm install -g propr-cli
propr setup

propr setup checks the host, signs in Claude Code in your browser, connects GitHub and starts the stack. Add the repository, and Coding agents shows Claude Code with its login state and default model. The quickstart covers the requirements.

Coding agents: Claude Code, Codex, Mistral Vibe, Antigravity and OpenCode, each with its health and default model — Coding agents after setup: each one with its login state and default model.

Pick the Claude model per issue

One label on the issue chooses the model. Configure the models once and the same IDs work in issue labels, plan settings, /switch and /review.

Model	Issue label	Use it for
Claude Opus 5.5 (default)	`llm-claude-opus55`	Planning, implementation and review
Claude Fable 5.1	`llm-claude-fable51`	The hardest planning, architecture and review work
Claude Sonnet 5.5	`llm-claude-sonnet55`	Routine implementation and PR follow-ups
Claude Haiku 5.5	`llm-claude-haiku55`	Fast, high-volume passes: summaries, triage, quick checks

Set the reasoning level globally, per model, or per issue with a level-* label, from low to max, including Claude Code's adaptive auto. Route Claude by phase as well: Fable for planning and review, Sonnet for routine implementation, with one branch and one history across all of them.

Claude Code in Agent Configuration: Claude Opus 5.5 as the default, Claude Fable 5.1, Claude Sonnet 5.5 and Claude Haiku 5.5, each with its 1M context and short alias, and the earlier models folded away — The Claude models configured in ProPR, each with its label and reasoning level.

Claude Code on pull requests

Comment on the pull request in plain language and Claude Code pushes the next commit to the same branch. /review fable51 gets a read-only critique from Fable with numbered findings and a score, /fix applies the findings you keep, and /switch hands future follow-ups to another model. Every comment command.

Put two model labels on one issue and ProPR opens one pull request per model, so Claude and Codex answer the same issue side by side. Longer objectives run as Goals, which Claude Code runs natively with your corrections reaching the session live. With Agent Tank on, ProPR shows the Claude capacity each task spent.

One pull request, six runs: the implementation, a review scored 6, the fix, a review scored 9, a branch update from main, and a follow-up from a PR comment — Every run on one pull request: the implementation, a review and the follow-ups.

The same path for Codex, Mistral Vibe, OpenCode and Antigravity CLI

Every supported agent runs through the same issue, workspace and pull request loop: the isolated run, the linked pull request, the follow-up commits and the task record with its prompts, logs and cost. Only the login and the model labels change, so switching agents mid-task keeps one history.

Codex

Runs the official Codex CLI on your ChatGPT login or an OpenAI API key. llm-codex-astra selects GPT-6 Astra, the default. Codex runs Goals natively too. Codex in ProPR.

Mistral Vibe

Signs in with a Mistral API key or your Vibe login. llm-vibe-mistral selects Mistral Medium 3.5, with GLM 5.3 alongside. Vibe does not run Goals. Mistral Vibe in ProPR.

OpenCode

Starts with free models, so an llm-opencode-big-pickle label costs nothing in tokens and gives comparison runs a zero-cost baseline. Bring any provider key for its paid models. OpenCode does not run Goals. OpenCode in ProPR.

Antigravity CLI

One Google login covers Gemini, Claude and GPT-OSS models. llm-antigravity-flash38 is the default, and llm-antigravity-argon selects Gemini 4 Argon, ProPR's recommendation for planning and review. Goals run as resumable sessions that take your corrections between steps. Antigravity CLI in ProPR.