RailCall Chat

Connect Chat to your station

Install or update the station, then start it. Keep that terminal open while using Chat.

curl -fsSL https://railcall.ai/install.sh | bash
railcall version
railcall studio

Download Chat from Chat downloads. On a Mac, choose Apple Silicon for an M-series chip or Intel for an Intel processor. Open Chat and connect to the local station. Use the station URL printed by the terminal if you need to open Studio.

Start with Auto, the recommended routing option. It chooses from the model routes available to your station under its policy. You still need a local model, an authorized collaborator, or a supported cloud key. See Models & Auto before enabling remote reasoning.

Start with an answer, then generate a PDF

First ask a short question to confirm that your chosen model responds. Then try a task with an explicit output and a clear boundary:

Using only the notes below, compare three rollout options for a
small support team. Explain assumptions, risks, and a recommendation.
Generate a PDF with a comparison table and a two-week rollout checklist.
Save it locally. Do not email or post it.

Notes:
- Four support agents handle 300 tickets each week.
- Start with drafting replies; every outbound reply needs approval.
- Pilot with one agent for one week before expanding.
- Track time saved, corrections, and failed runs. Do not invent results.

The model prepares the content; a PDF export action creates the file. Depending on installed capabilities, export can use the native action or a document-generation module. A written plan or a file name in a reply is not proof of export: open the actual artifact and check the text, tables, and assumptions.

For research tasks, explicitly request source links and inspect them. Web search and fetching pages are network operations. A generated PDF does not establish that its claims or calculations are correct.

Review actions before execution

For an integration request, inspect the editable action card: destination, content, required fields, and preview result. Missing credentials or inputs must be resolved before approval. An unconfigured preview is not a successful send. External sends require your approval; inspect the resulting status and receipt.

Save & run a workflow

In web chat, ask for a reusable research or reasoning workflow. A supported proposal appears as a review card: inspect the step instructions, choose Save workflow to my account, then Open saved workflow in Canvas. You can edit the connected nodes and save another version. Saving requires sign-in and does not execute the workflow or activate a schedule.

Chat proposals currently support linear research and reasoning steps. Hosted Canvas execution and direct web-to-local execution are not enabled by this feature. Use the existing Studio transfer to review a design locally; transferring it is not execution approval. Petiw desktop integration is planned, not available through this release.

In Studio Builder, compose a workflow and inspect each step. In the documented release, Save & run supports linear workflows whose steps are bound to commands. Missing bindings, required inputs, blocked previews, and unsupported conditional paths stop the runner. Complete setup and review before running again.

This is a manual run. It does not activate a recurring schedule or make a chat keyword into a persistent trigger. Use the station's workflow and scheduling controls where supported, and verify a real run before relying on unattended execution.

Optional hosted RailCall Assistant

Upload a file or generate a document in web chat

Sign in to web chat. Use Attach to choose one PDF, PNG or JPEG, then ask a question about it. Files may be up to 4 MB; PDFs may have up to five pages. Station extracts embedded PDF text first and uses English OCR for scanned pages and images. Check names and numbers: OCR can make mistakes. Relevant extracted text is sent to the selected cloud model.

To create a file, ask explicitly: Generate a PDF with a sourced summary of recent mathematics research, Create a DOCX summary of this document, or Create an XLSX spreadsheet from these notes. PatSUM prepares the content, a hosted RailCall Station creates the file, and the reply provides a private download. Generated files are small reports, limited to 64 KB; spreadsheet cells are literal values, not executable formulas.

Uploads and generated downloads expire after 24 hours. They belong to your account, not a public share link. Inspect the request trace for extraction, sources, actual cloud model, file execution and failures. Existing marketplace credit accounting applies; an unsuccessful run has no customer charge. A generated document is not proof that its claims or OCR text are correct.

This hosted path supports read-only search and page fetching. It does not grant access to your Chrome sessions or bypass Station browser policy. Interactive browsing, login, payment and other approval-gated actions remain governed by the existing Station browser and airlock.

Thinking modes in web chat

In web chat, choose Thinking beside the send controls before sending. Auto keeps the model defaults. Fast favors a concise answer, Balanced adds more checking, and Deep asks for more thorough analysis and may take longer. The choice stays fixed for that request; start a new request to change it.

Current-information questions can still search in Fast. In signed-in hosted chat, PatSUM chooses eligible cloud models according to the selected profile and provider availability. Modes do not grant extra tools or bypass marketplace credit checks. Native thinking controls vary by provider; inspect the trace for what actually ran. Deep shares the output limit and can run out of tokens.

Open Debug request to see the requested mode, actual model, applied control, host, timing and fallback. This shows execution events, not private model reasoning. If the modes are unavailable, the primary host needs a compatible update; Auto remains available.

RailCall Assistant is a separate, hosted preview option. It uses RailHub and a host station, such as Patrik's computer, with local models and potentially service-funded external reasoning. It is not the same as ordinary Auto on your own station.

Sign in with railcall market login, then reopen the model list. Availability depends on the service, worker health, queue, and usage limits; a visible entry does not guarantee a ready worker. Shared hardware has limited capacity, so requests may wait or be refused.

The preview admission window is set by the hosted service and shown on the downloads page and in the app; check there for the current date. It does not automatically convert to a paid subscription. Check the disclosure shown before submitting; read data boundaries and retention.

If setup, a model, or an export fails, use Troubleshooting. General crash recovery and arbitrary workflow resume are not promised by this release.