Open source · Self-hosted · All systems go
Free, open-source DORA and PR metrics for your GitHub repos. See which pull request is stuck on the pad, which pipeline is red and how long a fix takes to reach orbit. Run it yourself with one token; your data never leaves your server.
git clone https://github.com/vistik/Houston.git && cd Houston && composer setup
MIT licensed · Laravel + SQLite · launches from your laptop
Telemetry
Grouped the way the report reads, from how long a PR waits on the pad to how often production needs an abort. Every metric carries a plain-words note on how it is calculated.
A PR goes stale after 7 days without activity, and the threshold is a setting. Active developers is a headcount, not a ranking: there are no per-person tables.
Flight recorder · per repository
Houston reads a repo's pull requests, reviews, Actions runs and deployments, and splits each PR's flight into coding, pickup, review and merge delay. In working hours, so a PR opened on Friday evening doesn't look two days slower.
Also on the page: trends, the time-to-merge distribution, CI health and work in progress, for the last 7 days up to 12 months, or a custom range.
Mission log · before and after
Introduced AI code reviews, switched tools, set up a Slack bot that nudges reviewers: add it as an event and it shows as a line on every timeline. Click the line to compare the time since with the equally long stretch before it.
Add events under Settings → Events. The screenshot uses the demo data that ships with Houston.
Per team
A team is a group of people whose pull requests are reported on together, across every watched repository. Memberships have join and leave dates, so work counts for the team someone was on when they did it.
Set teams up under Settings → Teams and add people by their GitHub login.
How memberships count
B left halfway through, C joined later. Each person's PRs count for the team only while they were on it.
CAPCOM · MCP
Houston ships a read-only MCP server. Connect Claude Code or any MCP client and it becomes your CAPCOM: ask about cycle time, review load or a bad week in CI. The assistant reads the same reports you do, including the previous period to compare against.
php artisan mcp:start houston
/mcp with a bearer token; the endpoint stays off until you set one
Register the local server with Claude Code. Use the absolute path to your checkout.
claude mcp add houston -- php /path/to/Houston/artisan mcp:start houstonstdio · read-only · no token needed locally
Add this to your client's MCP config (stdio).
{
"mcpServers": {
"houston": {
"command": "php",
"args": ["/path/to/Houston/artisan", "mcp:start", "houston"]
}
}
}1. On the server, set a token in .env. Any long random string works.
HOUSTON_MCP_TOKEN=[a long random string]2. Point your client at /mcp and send it as a bearer token.
claude mcp add --transport http houston http://localhost:8000/mcp \
--header "Authorization: Bearer $HOUSTON_MCP_TOKEN"No token set: /mcp answers 404 · wrong token: 401
● calls
get-velocity-report
read-only
{ "repository": "laravel/framework", "days": 30 }list-repositories toolget-velocity-report toollist-pull-requests toolgive-feedback promptThe big board
A hands-off, auto-rotating view meant for a screen on the wall, like the front screens in mission control. It is always dark, whatever theme the app uses, and shows one repository at a time.
?rotate=/wall for every repo, /repos/{owner}/{name}/wall for oneGo / no-go
Set targets for any metric and Houston rates it as elite, good, medium or low wherever it appears. Counts of work are benchmarked per week, so one target holds for every reporting period.
It is opt-in: nothing is rated until you set targets. When you want a starting point, 31 standards are one click away: 4 from DORA, 4 from LinearB and 23 suggested targets.
| Metric | Elite | Good | Medium | Low |
|---|---|---|---|---|
| Deploy frequency | 1/day | 1/week | 1/month | < 1/month |
| Lead time to production | 1 d | 7 d | 30 d | > 30 d |
| Change failure rate | 5% | 10% | 15% | > 15% |
| Time to restore | 1 h | 1 d | 7 d | > 7 d |
Thresholds a value must meet for each tier; anything worse than medium rates low.
Flight rules
A metric is only useful if the team believes it. These are the flight rules, and most of them are settings.
Durations that measure people (pickup, review, time to merge, recovery) count working time, so a PR opened on Friday evening doesn't look two days slower than one opened on Monday. Working days and hours are settings, and they count in the timezone you pick for your reports; the default is Monday to Friday, UTC. CI durations stay wall-clock.
Pull requests from Dependabot, Renovate and other bots are left out of the metrics, so they don't skew people numbers. The bot backlog is reported on its own.
Each number sits next to the same metric over the equally long period before it. "Last 30 days" is today and the 29 days before it, from midnight in your timezone, so every bar on a chart is a whole day and the bars add up to the number. Metrics that sum up one value per PR show the median and the average, so a skewed spread shows.
There are no author or reviewer tables. Houston reports on teams and repositories, never on individual people.
A deployment counts as failed when the next one ships a revert, hotfix or rollback PR, matched on title, branch or labels. The pattern is configurable. There is no incident data behind it.
GitHub Deployments to an environment named like prod, production or live. Repos without them fall back to published GitHub Releases.
Launch sequence
git clone https://github.com/vistik/Houston.git
cd Houston
composer setupcomposer setup installs the PHP and Node dependencies, creates .env, generates the app key, runs the migrations and builds the frontend.
gh auth loginWith GITHUB_TOKEN empty, Houston uses the GitHub CLI's token. On a server, set GITHUB_TOKEN in .env instead.
composer devStarts the scheduler, five queue workers, a log tail and Vite. It doesn't start a web server: serve the app with Laravel Herd, or run php artisan serve next to it. Then click Watch repository and enter owner/name.
Syncs run on the queue, and the scheduler refreshes every watched repo every 15 minutes. On a long-running server, run a queue worker and the scheduler.
Houston asks for a login out of the box; php artisan houston:install creates it. Turn it off on your laptop or a private network if you like; keep it on for a public server.
Yes. Houston is MIT licensed. There is no paid tier, no hosted version and no seat count. No launch fees either.
On your server. Houston keeps a local cache of pull requests, reviews, workflow runs and deployments in SQLite by default, or in any database Laravel supports. Nothing is sent anywhere but GitHub's API.
Authorized personnel only: whoever has the login. Houston has one login, on by default: php artisan houston:install creates it, and you can turn it off on your own machine. There are no per-user accounts. The /mcp endpoint has its own token.
No. Houston reports on teams and repositories, never on individual people.
No. Houston reads GitHub only: pull requests, reviews, Actions workflow runs, and Deployments or Releases.
A deployment counts as a failed change when the next deployment ships a revert, hotfix or rollback pull request, matched on its title, branch or labels. The pattern is configurable. It is inferred from pull requests, not from incident data.
It syncs every 15 minutes and pauses when fewer than 50 requests are left in the current window, leaving room for anything else that shares the token. Paused syncs resume when the window resets.
Point it at your repos and see your own numbers. Free and MIT licensed. One GitHub token, one SQLite file, nothing to sign up for.