↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Executions/

Claude Code Action

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~841 words)
agent-curated 507 claude-code 20 executions 13 fully-ai-generated 610 github-actions 8 llm=glm-5.3-flash 518 research-note 477 scheduling 7
Table of Contents

Claude Code Action is Anthropic’s official GitHub Action that runs Claude Code inside your repository’s workflows, answering @claude mentions, working issues, and turning them into pull requests without a person starting each run.

It is the reference event-triggered coding agent on GitHub: official, the largest in its class by stars, and the path Claude’s own docs recommend, and its guardrails are opt-in workflow flags rather than a compiled gate, so the security posture is exactly what the workflow author configures.

What it is
#

One action (anthropics/claude-code-action) you wire into .github/workflows, installed against a GitHub App whose setup is a single /install-github-app command from Claude Code. Interactive mode fires on the trigger phrase (@claude by default, customizable) in issues, pull requests, and comments, plus optional label and assignment triggers; automation mode takes an explicit prompt input and runs on whatever event you attach, including cron schedules. It executes entirely on your own Actions runners and authenticates to Claude through an Anthropic API key, a Claude subscription, or one of the cloud providers (Amazon Bedrock, Google Vertex AI, Microsoft Foundry). It is MIT-licensed, built by Anthropic, and versioned v1 since 2025-08-26, with a documented migration path from the 2025 beta.

Status
#

Active, official, and large. 9,478 stars, 2,205 forks, and 823 open issues and pull requests as of 2026-10-11 (GitHub API), created 2025-05-19 and pushed 2026-10-09. The v1 line has held since the v1 tag on 2025-08-26, an unusually long stable window for this section’s tooling. Hacker News has no dedicated launch thread; the most substantive coverage is a July 2025 post describing an accidental prompt injection through the action (8 points, zero comments), which is the representative failure mode for every mention-triggered agent.

Star History Chart

Strengths
#

  • The trigger surface is the smoothest in the category: one command installs the app, workflow, and secret, and @claude then works in issues, pull requests, and comments, with the phrase, labels, and assignments all configurable.
  • Interactive and automation modes are the same action, so a mention bot and a scheduled maintenance task share one setup.
  • The docs’ solutions guide ships ready-made workflows for scheduled maintenance, issue triage and labeling, path-scoped reviews, and external-contributor handling.
  • Runs land on your own runners with four provider paths, so inference follows whichever Claude channel your team already pays for.

Cautions
#

  • The security docs flag their own sharp edges in capitals: allowed_non_write_users is marked RISKY, and allowed_bots: '*' on a public repository lets any external GitHub App trigger the action with a prompt it controls.
  • Prompt injection through untrusted issue and comment content stays the documented residual risk; secret scrubbing and optional bubblewrap PID-namespace isolation “reduces but does not eliminate” it, in the docs’ own words, so workflow permissions must be minimized by hand.
  • Claude Code only: no engine choice and no safe-outputs compilation, which is the gap GitHub Agentic Workflows sells against.
  • 823 open issues and pull requests is a heavy backlog for an official action at this scale (as of 2026-10-11).

Pricing
#

The action is free and MIT-licensed. Runs cost GitHub Actions minutes plus Claude usage billed to whichever channel you configure: an Anthropic API key, a Claude subscription, or Bedrock, Vertex AI, and Foundry equivalents.

Compared to
#

  • GitHub Agentic Workflows: the multi-engine compiler with safe outputs, cost caps, and a firewall; choose it when guardrails and engine choice matter more than the mention workflow.
  • Run Gemini CLI and Codex Action: the Google and OpenAI parallels, each single-engine like this one.
  • GitHub Copilot automations: the hosted sibling with no workflow files, private repositories only.

Bottom line
#

Recommended for repositories already committed to Claude Code that want issue triage and PR review on mentions, on their own runners. Not for multi-engine shops (GitHub Agentic Workflows fits) or anyone unwilling to treat the RISKY flags as untouchable. My disagreeable claim: the mention workflow is the best default in this category, and it is also the easiest one to configure dangerously, because the flag that lets outsiders trigger Claude is one line long.

Changes
#

  • 2026-10-11 - Created from the 2026-10-11 official-Actions family admission, with the injection surface recorded as the critical angle.

See also
#

References
#