AI GitHub Issue Triage with FixClaw: Human in the Loop
Open source maintainers do not burn out from writing code. They burn out from the inbox. A hundred open issues, half of them duplicates, a third missing a stack trace, and somewhere in there the one real bug that will bite a paying user tomorrow. FixClaw is an AI GitHub issue triage tool built to take that load off without taking the keys away from you.
This is a product landing page, not a tutorial. If you want to build your own triage system from scratch, skip to the Related Reading at the bottom. If you want github issue automation that reads like a helpful junior maintainer and waits for your approval before posting anything, keep reading.
The problem every maintainer knows
You open GitHub on Monday. Forty new issues since Friday. Eight of them are the same crash on Windows, filed by eight different people who did not search before opening. Four are feature requests disguised as bugs. Two are one-line “it does not work” reports with no version, no OS, no trace. One is a real security report buried between a typo and a theming request.
Triage takes hours. You copy-paste the same “can you share your Node version and a minimal repro” comment five times. You add the same needs-info label ten times. You close duplicates and link them back. By the time you finish, your focus time for the week is gone.
The usual answers do not fix this. GitHub Actions with regex rules catch the obvious cases and miss the rest. Generic issue bots auto-reply with canned text that annoys reporters. Paying a part-time triager works until volume spikes. None of them read the issue the way you would.
What FixClaw actually does
FixClaw is ai issue management for GitHub repos. One sentence: it reads every new issue, finds duplicates in your existing history, classifies severity and type, suggests labels, drafts a triage comment, and waits for you to accept or edit.
It installs as a GitHub App. You point it at a repo, optionally set a few rules, and from that point on every new issue gets a structured triage suggestion within minutes. You stay the maintainer. FixClaw stays the assistant.
How it fits into your workflow
The flow on every new issue looks like this.
new issue opened
> FixClaw reads issue body, comments, linked code and docs
> pulls last 200 issues from the repo for duplicate context
> Claude classifies: bug / feature / question / duplicate
> generates labels, links related issues, drafts a response
> posts a triage suggestion as a comment on the issue
> you accept, edit, or ignore
No auto-reply goes to the reporter unless you set a rule for it. The default mode is suggest, do not act. You see the proposed labels, the proposed duplicate link, the proposed response, and one click applies them.
If you want more automation, you turn it on per label. For example: auto-apply label=needs-info when confidence is above 0.9, auto-close as duplicate when confidence is above 0.95 and the source issue is less than 30 days old. Rules are per repo, not global.
Why Claude under the hood
FixClaw uses Claude because three of its properties matter for triage specifically.
Long context. Duplicate detection is the hardest triage task. A generic classifier sees the new issue and maybe top-k nearest neighbors from a vector index. Claude sees the new issue plus the full text of fifty candidate issues plus the linked code snippets, all in one prompt. That is the difference between “these look similar” and “this is actually the same bug, filed three weeks ago, and here is the line that proves it.”
Tool use for structured output. Triage is a structured action: a list of labels, a severity, a duplicate link, a draft comment. FixClaw defines these as tool schemas and Claude returns them as validated JSON. No regex parsing of prose, no hallucinated label names that do not exist in your repo.
Reasoning. For hard calls, FixClaw turns on extended thinking. “Is this a duplicate of #1247, or a genuinely new case with overlapping symptoms” is the kind of question that benefits from a model that can actually think through the difference before answering.
What sets FixClaw apart from auto-responders
Generic issue bots paste the same template reply into every thread. Reporters stop reading them after the second one. FixClaw is grounded in your repo, not in a generic model of how issues look.
It reads your actual issue history, not a prebuilt taxonomy. If your team uses area/parser as a label, that is what gets suggested. If you have a convention where P0 means “prod is down” and P3 means “nice to have”, FixClaw learns that from the last hundred closed issues.
It reads docs and code references linked in the issue. When a reporter pastes a stack trace pointing at src/auth/session.ts:42, FixClaw pulls that file and includes it in the triage context. The suggested comment can reference the actual line.
Every suggestion carries a confidence score and a one-line “why”. You see that the duplicate suggestion is 0.87 confident and that the reasoning is “same error message, same OS, same version, different user”. You decide if that is enough.
The default is human in the loop. Nothing posts, labels, or closes without your rule or your click. If you want a fully autonomous bot that runs your repo, FixClaw is the wrong tool. If you want a triage co-pilot, it is the right one.
Typical use cases
Small to medium open source projects. You maintain a library with 20 to 100 open issues and you are one person or a team of three. Triage is eating your weekend. FixClaw cuts the time per issue from minutes to seconds because the first draft is already there.
Enterprise repos with dedicated triage. A large internal monorepo where triage is someone’s full-time job. FixClaw takes the first pass so the human triager reviews suggestions instead of starting from zero. The same team can cover double the volume.
Community-heavy projects. Projects where faster first response keeps contributors engaged. FixClaw can post a polite, accurate “here is what we need from you” within minutes of an issue opening, instead of three days later when a maintainer gets to it.
What FixClaw doesn’t do
It does not write your code. Issue to PR generation is a different product category, and mixing it in would break the human-in-the-loop promise. If you want code written, use a coding agent and keep triage separate.
It does not replace maintainer judgment on feature decisions. FixClaw will label a request as feature and suggest a draft acknowledgment. It will not decide whether you accept the feature. That stays yours.
It does not touch pull requests. Issues only. PR review is a different problem with different tradeoffs and we do not want to half-solve it.
It does not ship canned replies. Every draft comment is generated fresh from the specific issue. No template spam.
Setup in five minutes
- Install the GitHub App on your account or organization.
- Pick the repos you want FixClaw active on.
- Optional: open the dashboard and set per-label rules (auto-apply, auto-close, confidence thresholds).
- Open a new issue or wait for one. First triage suggestion appears as a comment within minutes.
No YAML in your repo. No workflow files to maintain. No secrets to rotate. The App holds the permissions, FixClaw talks to GitHub and Claude, you review the results on the issue itself.
Pricing and model
Free tier for small repos, under 50 issues per month. Paid tier for higher volume, priced per issue or per month. Open source projects with an active community typically qualify for free usage with a light integration credit in the issue template.
Enterprise self-host is available for teams who need the pipeline running inside their own infrastructure. Same product, your VPC, your Anthropic API key.
Check fixclaw.com for current numbers, they move as the product matures.
Data handling and privacy
FixClaw reads the text of issues and their comments. For public repos, that is public data. For private repos, the data flows through the Anthropic API under a standard data processing agreement, with no training on your content and the option to run in an EU region if you need it.
Nothing is stored beyond what FixClaw needs to operate: the repo settings, rule configuration, and the suggestions themselves so you can audit them later. Issue bodies are not retained after the triage pass completes.
If you self-host, none of this leaves your infrastructure. You bring your own Anthropic key and the data path is yours end to end.
How it was built
FixClaw is built on the Claude API with tool use as the core mechanism for structured triage actions. The backend is TypeScript, webhook triggered, stateless between events. A GitHub webhook fires on issues.opened, FixClaw pulls repo context, calls Claude with the triage tool schema, and posts the result back as a comment.
The duplicate detection layer uses a mix of embeddings for the initial shortlist and Claude with long context for the final decision. Cheap recall, expensive precision, only where it matters. Extended thinking is enabled selectively when the shortlist is ambiguous, which keeps average cost per triage predictable.
The architecture draws on the same patterns used across production agent systems: structured tool calls over free-form generation, prompt caching on the repo-context prefix so common reads stay cheap, strict JSON validation on every tool response. For the deeper version of that architecture, the production AI agent architecture guide covers the full pattern.
If you want to build something like this yourself, the build an MCP server in TypeScript tutorial and the Claude Code SDK agents guide cover the moving parts. The reasoning behind using AI for triage in the first place is written up in GitHub issue management with AI.
When FixClaw is the right tool
You maintain between one and fifty active repos. Your issue volume is above twenty per month, enough that manual triage actually costs you time. You want triage help, not a bot that decides on your behalf. You want labels, duplicate links, and a first-draft reply waiting for you, and you want to keep the final click.
When it isn’t
Your triage is fully deterministic and a GitHub Action with labelers and regex rules handles it. Buying an AI tool on top would add complexity without value.
Your issues are deeply domain specific and only a subject matter expert can classify them. FixClaw will be wrong often enough to annoy you.
You want a fully autonomous bot that runs the whole repo. FixClaw is human in the loop by design and will stay that way. If autonomy is the goal, a different tool fits better.
Related Reading
- GitHub issue management with AI for the broader case
- Production AI agent architecture for the patterns FixClaw uses
- Build an MCP server in TypeScript and Claude Code SDK agents if you want to build your own
Maintainers burn out because of triage, not because of coding. FixClaw takes the triage load off while keeping you in control of every label, every link, every close. Start with one repo. See the first suggestion. Decide from there.