AI agents · MCP
Let your AI agent ship and run your apps.
Railyard has a built-in MCP server. Claude, Claude Code, Cursor or any MCP client can deploy, roll back, read logs and metrics, query your database and manage domains on your own servers. You choose its access, and anything that removes or overwrites data waits for a person.
MCP server
42 tools on one endpoint.
POST https://app.railyard.run/api/mcp speaks the Model Context Protocol, the current revision and the earlier ones, over streamable HTTP. Claude and other clients sign in with OAuth on a Railyard consent page; code clients can send a Railyard API token instead. It sees only the team you authorize.
| Tool | What it does | Needs |
|---|---|---|
| Account and templates | ||
| whoami | The team, its plan and limits, and this connection's access | read |
| list_templates | Search the one-click templates and stacks | read |
| launch_template | Create an app or stack from a template and deploy it | write |
| Apps | ||
| list_apps | Every app with status, URL, server and environment | read |
| get_app | One app: status, repository, domains, processes, variable names and recent deploys | read |
| create_app | Create an app from a Git repository and start its first deploy | write |
| create_environment | Copy an app into a new environment, such as staging | write |
| clone_app | Copy an app's settings and variables into a new app | write |
| promote_app | Promote an environment to the app it came from and deploy the same commit | write |
| scale_app | Set how many containers a worker runs | deploy |
| set_maintenance | Turn maintenance mode on or off | deploy |
| delete_app | Delete an app and its data | write + approval |
| Deploys | ||
| deploy_app | Deploy the latest commit, or an exact commit, tag or branch | deploy |
| list_deploys | Recent deploys, filtered by app or status | read |
| get_deploy | One deploy: commit, trigger, timing and the error if it failed | read |
| deploy_logs | The build and release log of a deploy | read |
| rollback_app | Go back to an earlier successful deploy, with its variables | deploy |
| cancel_deploy | Cancel a queued, scheduled or running deploy | deploy |
| Runtime | ||
| app_logs | Runtime logs, filtered by text, level, process or time range | read |
| list_processes | Running containers with health, restarts, CPU and memory | read |
| restart_app | Restart an app's containers, or one process's | deploy |
| run_command | Run a one-off command in the app's container | write |
| get_command | Status, exit code and output of that command | read |
| list_scheduled_tasks | Cron tasks with their schedule and last run | read |
| run_scheduled_task | Run a scheduled task now | write |
| Variables | ||
| get_env | Variables with their values; sealed ones stay hidden | write |
| set_env | Set a variable, applied on the next deploy | write |
| unset_env | Remove a variable | write + approval |
| Domains | ||
| list_domains | Custom domains with DNS, verification and certificate state | read |
| add_domain | Add a domain and get the DNS records to create | write |
| remove_domain | Remove a domain | write + approval |
| Databases | ||
| list_backups | Recent backups of the app's database | read |
| create_backup | Back up the database now | write |
| restore_backup | Restore the database from a backup | write + approval |
| list_tables | Tables with row counts and sizes | write |
| query_database | One read-only SQL query, up to 200 rows | write |
| Servers | ||
| list_servers | Servers with status, provider, load and pending security updates | read |
| get_server | One server in detail, with the apps it runs | read |
| reboot_server | Reboot a server | write + approval |
| Monitoring | ||
| app_traffic | Requests, error rate, latency and busiest paths | read |
| list_alerts | Metric alerts and recent incidents | read |
| list_activity | The audit trail: who changed what, and when | read |
Ask it
Plain requests, real answers.
- Why did the last deploy of my API fail? Fix it and redeploy.
- Roll the shop back to yesterday's release.
- Which apps had 5xx errors in the last hour?
- Spin up a staging copy of the web app on the main branch.
- How many signups did we get this week? Query the database.
- Add shop.example.com to the storefront and tell me which DNS records to create.
Three built-in prompts
Ready-made prompts for the common jobs.
- debug_failed_deploy
- Finds why an app's latest deploy failed and suggests the fix.
- app_health_report
- Status, traffic, errors, processes and alerts for one app.
- fleet_overview
- Every server and app on the team, with what needs attention.
Connect
Add it to your agent.
Claude finds Railyard in its connectors directory. Code clients take the endpoint URL and sign in through your browser; if a client can't do OAuth, create an API token in Settings → Access & tokens and send it as rly_....
Customize → Connectors Search "Railyard" → Connect Approve access on the Railyard page
claude mcp add --transport http railyard https://app.railyard.run/api/mcp
{
"mcpServers": {
"railyard": {
"url": "https://app.railyard.run/api/mcp"
}
}
} Guard rails
Scoped access, human approvals.
- Read access lists and inspects apps, servers, deploys, logs, traffic and activity. Nothing else.
- Deploy access also deploys, rolls back, restarts, scales and toggles maintenance mode.
- Write access also changes variables, domains, backups and templates, runs commands and reads data with read-only SQL.
- Deleting an app, removing a variable or domain, restoring a backup and rebooting a server always wait for an admin to approve them in Railyard.
- Sealed variables stay masked and can't be changed by an agent, and every call is in the activity log, marked as coming from MCP.
> Deploy storefront and check that it's healthy
→ list_apps
storefront · running · bold-moose
→ deploy_app storefront
deploy #42 succeeded · health check passed
→ app_logs storefront
100 lines, no errors
> Remove the STRIPE_TEST_KEY variable
→ unset_env storefront STRIPE_TEST_KEY
waiting for a human to approve it in Railyard
Shell and HTTP
The CLI and REST API work too.
Agents that run shell commands can use the same CLI you do, and anything can call the REST API with the same tokens. API errors exit non-zero and railyard run exits with the remote command's own code, so an agent knows when something failed.
railyard login https://app.railyard.run railyard deploy shop --follow railyard logs shop --follow --process web railyard run shop -- bin/rails db:migrate railyard rollback shop 218
Questions
MCP on Railyard, answered.
Does Railyard have an MCP server?
Yes. It's built in at /api/mcp on every Railyard installation and listed in Claude's connectors directory. It works with Claude, Claude Code, Cursor and any client that speaks MCP over streamable HTTP.
How do I connect Claude to Railyard?
In Claude, open Customize → Connectors, search for Railyard and choose Connect, then approve access on the Railyard page that opens. In Claude Code, run claude mcp add with the endpoint URL.
Can an AI agent delete my app or my data?
Not on its own. Deleting an app, removing a variable or a domain, restoring a backup and rebooting a server create an approval request, and nothing changes until a team admin approves it.
Does the agent see my secrets?
Read and deploy access see variable names only. Values need write access, and sealed variables stay masked. Write access can also run commands inside your app's container, so give it only to agents you trust.
Can it query my database?
With write access it can run one read-only SQL query at a time. The query runs in a read-only transaction with a time limit and returns at most 200 rows.
Your own servers
Agent experiments run on servers you already pay for.
Agents deploy often and try things. On a platform that bills per service or per usage, every experiment is a line on the invoice. On Railyard, an agent's staging copies and retries are containers on servers you already pay for. Its staging copies and the data they hold stay on your own servers, and you can cut its access in one click.
Ready to own your production infrastructure?
Start your 14-day trial or book a 48-hour migration call.
14-day trial on every plan. One flat price per team, your servers at your provider's price.