
Mergewarden
io.github.sjh9714Published by GitHub account sjh9714v0.9.0Updated Oct 24, 2026
Check whether a coding agent's changes stayed inside the scope it was given.
Overview
MergeWarden checks whether a coding agent's changes stayed within their assigned scope, and flags pull requests that need context before review.
- What you can do
- Check whether a coding agent's changes stayed inside the scope it was given, using deterministic rules rather than a language model. Flag pull requests missing linked issues, with thin descriptions, skipped templates, or oversized changes. Run a detailed risk scan on a pull request covering workflow permissions, agent instruction files, untrusted prompt inputs and install scripts. Report incomplete evidence as incomplete rather than as a pass.
- When to use it
- Useful when reviewing pull requests from coding agents or outside contributors and you want a quick, rule-based check of scope and review readiness before reading code. Also suited to spotting risky changes to CI workflows, agent instruction files or package install scripts.
- Requirements
- Runs as a local npm package over stdio. The README's CLI examples use a GH_TOKEN environment variable for a larger authenticated queue; the manifest declares no authentication or environment variables. Network access to the public GitHub API is needed.
Installation
In SourceWeft
- Open Mergewarden in the dashboard and add it to a workspace.
- Enable the server for the chats that should use its tools.
Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.
Other MCP clients
Follow the launch instructions in the repository.
README
MergeWarden
[Release] [CI] [MergeWarden] [License]
See which pull requests need context before code review.
Paste a public GitHub repository. MergeWarden surfaces missing issue links, thin descriptions, skipped templates, and oversized changes. No login. No AI.
This is an experimental tool. Review-time savings and repeat maintainer use have not been demonstrated. Product expansion is on hold; see the roadmap and evidence limits.
Review a public repository
The getting started guide explains queue states, detailed PR results, and the Action install path.
Paste owner/repository or a full GitHub repository URL. The browser reads the
latest thirty open pull request summaries, removes trusted repository roles,
base repository branches, and known maintenance automation, then loads details
for at most ten external pull requests.
Rows are ordered by deterministic facts a maintainer would otherwise check before reading code.
First contribution is shown as context and never used as a score. MergeWarden does not call a contribution spam, low quality, or AI generated. It never closes, labels, scores, or comments on a pull request.
The queue uses public metadata and the base branch pull request template. It does not fetch changed file contents, execute code, use a backend, or store the target.
For a larger authenticated queue, use the CLI.
Run the detailed PR risk scan
The same page accepts a full pull request URL or owner/repository#number.
Every queue row links to this detailed scan.
The detailed scan checks four security boundaries.
The configuration reference lists every deterministic rule and severity.
Run the same scan from a terminal without cloning the target repository.
Add the Action
The Action automates the detailed PR risk check. It does not order the review queue.
Create .github/workflows/mergewarden.yml.
comment: auto stays quiet when there is nothing actionable and updates one
comment when a finding needs attention. Existing Action defaults are unchanged.
For an immutable install, pin the release commit.
MergeWarden does not publish or recommend a mutable v0 tag.
Safety boundaries
- The web app talks directly to the public GitHub API and has no telemetry.
- The queue reads metadata and templates at each PR's exact base commit. The detailed scan reads only the files required by deterministic rules.
- No checkout. MergeWarden does not execute pull-request code in the web app or Action.
- Policy comes from the exact base commit, never the untrusted PR head.
- Analysis never calls a language model.
- Incomplete evidence is reported as incomplete and never presented as a pass.
- Browser sample limits are explicit: reaching 30 summaries or leaving external PRs beyond the ten-detail limit does not produce a complete queue.
Read the security model and evidence model for the full trust boundary.
Evidence and advanced interfaces
Does triage help? records the existing queue measurement and its main limit. It measured whether rows discriminate, not whether maintainers save time. The current web queue remains a product experiment until maintainers confirm that value.
The documentation index includes configuration, Action and CLI references, coding-tool integrations, reproducible studies, and the MCP server.
Contributing
Every new rule needs a passing fixture, a failing fixture, and a report snapshot. See the contribution guide.
License
Source: README.md at commit aba7262
Tools
0Version history
1- v0.9.0LatestSep 16, 2026
Score
50% of the applicable checks met
How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.
- Reached in the last 30 days25 pts
Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.
- Auth declared10 pts
Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.
- Tools listed15 pts
Answer tools/list without a credential, or ship a package that starts in a sandbox without one.
- DescriptionRequired15 pts
- Logo10 pts
- Documentation15 pts
- License10 pts