--- title: "Railway Agent, IPv6 Support" date: 2026-04-24 number: 0287 url: https://railway.com/changelog/2026-04-24-railway-agent --- # Railway Agent, IPv6 Support If an AI can tell you why your deploy is red, it should be able to make it green. This week, the Railway Agent is live for every workspace. It can create services, connect databases, walk you through a red deploy, and open a pull request with the fix, all from a chat panel in the Railway dashboard. Also, outbound IPv6 egress is now available on every service too, no feature flag required. Let's get into it! 🚄 ## Railway Agent The Railway Agent is a chat-based AI assistant built into the Railway dashboard (and [as of last week](https://railway.com/changelog/2026-04-17-remote-mcp/#railway-agent-in-the-cli), the CLI) that can operate the platform on your behalf. It has access to the same primitives you do, so it can act across your entire project: create services, set variables, connect databases, wire up networking, adjust resource limits, read logs and metrics, and walk through a failing deployment. [Video: Ask the Agent to create a service, diagnose a failing deploy, or open a PR with the fix] [Smart Diagnosis](https://railway.com/changelog/2026-02-27-smart-diagnosis/#smart-diagnosis-to-priority-boarding) folds into the Agent too. Ask it to investigate a failed deploy and it reads the build and runtime logs, correlates them with your service configuration and recent code changes, and produces a short explanation of what went wrong. If the fix lives in your code, it opens a pull request against your repository so you can review and merge. **Pricing** Agent usage is billed at the exact per-token rates for the underlying LLM, with no Railway markup. Under the hood, the Agent is powered by a mix of Anthropic models and routes each request to the smallest one that can handle it, so costs stay proportional to the difficulty of the task. Usage draws from the same included credit as the rest of your plan, so the $5 on Hobby and $20 on Pro that already covers your compute also covers your Agent. You're only charged beyond that once your combined compute and Agent spend exceeds what your subscription includes. Total Agent usage appears as a separate line item on your invoice. Of course, we want to make sure you never get a surprise bill. That’s why every workspace ships with a default Agent hard limit ($5 on Hobby, $20 on Pro), and you can define custom limits in your [Workspace Usage page](https://railway.com/workspace/usage). [Image: Set separate hard limits and email alerts for Compute and Agent usage] Try it out and let us know what you think [on Central Station](https://station.railway.com/new?type=feedback). ## IPv6 support [Video: Enable outbound IPv6 in your service’s network settings] Outbound IPv6 egress is now available to every workspace! If you need to connect to external services that only expose IPv6 addresses, your Railway services can now reach them. To enable it, open your service's Settings tab, scroll to the Networking section, and toggle Enable Outbound IPv6. Redeploy the service for the change to take effect. Outbound IPv6 is disabled by default. Enabling it does not affect your service's existing IPv4 outbound connectivity, both work concurrently when the toggle is on. While this setting is disabled, IPv6 connection attempts will fail with "Network is unreachable" or `ENETUNREACH`. As always, let us know what you think on [Central Station](https://station.railway.com/new?type=feedback). ## Fixes and improvements - We fixed Railway-purchased custom domains stalling during setup when the zone had a conflicting record at the apex. The setup workflow now clears same-host A, AAAA, and CNAME records before writing the ANAME, so cert issuance no longer gets stuck at `VALIDATING_OWNERSHIP`. We also stopped the service networking UI from showing "Waiting for DNS to propagate..." once the cert is already valid. - We fixed regenerating a MySQL database password only rotating one of the user's host variants, leaving socket logins (`root@localhost`) out of sync with the TCP user (`root@%`). The reset now updates both host variants atomically in a single `ALTER USER` statement. - We fixed scaling a cron service via the API starting an extra container that ran the cron's start command off-schedule. Scale workflows no-op on cron services now, since the schedule is the only legitimate trigger for a cron container. - We fixed quoted phrases in the log filter bar getting silently unwrapped by the router, which caused the filter parser to treat the space as an implicit AND (typing `"WARN level"` displayed as `WARN AND level`). Search-param round-tripping no longer strips the quotes. - We fixed WebSocket connections being counted as status 0 GETs in a service's HTTP metrics, which skewed request counts and duration percentiles. WebSocket traffic is excluded from request count, status breakdown, and latency queries now. - We fixed the deploy image input rejecting Google Artifact Registry images from any region other than `us-west1`. All `-docker.pkg.dev` hosts are accepted now, so `asia-southeast1`, `europe-west1`, and the rest deploy as expected.