RailCall is a free local AI runtime for developers. Nine of every ten requests are answered on your own machine for $0, at the quality you would get anyway. Only the hard tenth leaves to a bigger model, metered against a budget you set. Type what you want done and build it in Studio.
MIT · log in only for the hub and cloud agents · use the full power of the compute you already own
This chat starts on RailCall’s servers. Connect your own computer from the workspace when you’re ready. Default models are chosen for your machine: the stronger it is, the more runs locally.
Set up local compute$0 RailCall fees for local workflow flows.
Human approval for protected actions.
Signed receipts of recorded actions.
Run saved workflows on your own machine.
One hundred requests arriving. The dashed line is every request sent to Fable 5.1 / GPT-6 Astra at list. The pink line is RailCall doing the same work at the same quality: flat at $0 while your machine answers, a step only when a request overflows. Drag the share.
90% of the model cost is gone because nine of ten requests never left your machine. The metered tenth is priced as a usage fee against the budget you set, so the bill lands 80% below the top model. 90% local is the design target, not a promise.
Type what you want. Studio composes it from a library of 1,000+ nodes: the CRMs, inboxes, calendars, databases, file stores and chat tools your team already uses, plus any MCP tool. It runs on your machines, the hard step overflows only when it must, and every send stops at the airlock.
One workflow, 30 nodes. Four triggers feed it: a web form, the inbox, the help desk and chat. Seven agents do the work on your machines for $0: triage, research, pricing, support, QA, outreach, reporting. One hard clause overflows to a bigger model and is metered. Four humans decide: a manager, finance above $100K, a supervisor on support, and you on the morning brief. Then the CRM, the e-sign, the replies, the refund and the brief fire, each with its own receipt. Agents are nodes too: drop one into any flow.
Each receipt is signed with Ed25519 by the machine that did the work and linked by hash to the one before it. Click a row to read the receipt, then try to alter it: the signature fails and the chain breaks. Anyone can verify the whole ledger offline with one command, no account, no server.
| # | step | ran on | receipt · chain | verify |
|---|---|---|---|---|
| 1278 09:41:00 | New lead arrives web form · lena@harbor.co · local · $0 | mac-studio-02 | 096882f1… ← 00000000… | ✓ Ed25519 · chain ok |
| 1279 09:43:17 | Research agent · enrich agent node · local · $0 | mac-studio-02 | 7af8ed6f… ← 096882f1… | ✓ Ed25519 · chain ok |
| 1280 09:45:34 | Outreach agent · draft intro agent node · local · $0 | laptop-sara | 476ec3c8… ← 7af8ed6f… | ✓ Ed25519 · chain ok |
| 1281 09:47:51 | Airlock · intro approved manager step · human | laptop-sara | 48b38495… ← 476ec3c8… | ✓ Ed25519 · chain ok |
| 1282 09:49:08 | CRM · create deal Northwind · $38,000 · local · $0 | mac-studio-02 | 7f484eb5… ← 48b38495… | ✓ Ed25519 · chain ok |
| 1283 09:51:25 | Email · send intro to lena@harbor.co · local · $0 | mac-studio-02 | cf408d52… ← 7f484eb5… | ✓ Ed25519 · chain ok |
| 1284 09:53:42 | Contract clause review hard step · overflow · metered | mac-studio-02 | 684ceb1a… ← cf408d52… | ✓ Ed25519 · chain ok |
Example ledger for the flow above; the format, the signature scheme and the railcall verify check are real. Receipts never leave your machines unless you put them on the hub.
An agent watches, decides on your machine, acts through the airlock and leaves a receipt, then does it again. Your own agents are free and unlimited. Agents we host run 24/7 for $5 each.
Describe the CRM your team needs and Studio builds the app from nodes: contacts, pipeline, tasks, inbox, reports. It runs on your machine with your data in a file you own, and you change it by typing in the chat rail.
Every app Studio builds carries the same guarantees as a workflow: local by default, overflow metered against your budget, writes through the airlock, receipts anyone can verify offline.
RailCall picks the local model package for your machine. The bigger the machine, the bigger the package, the more answers never leave it, and the smaller the bill. The package keeps improving: as better local models ship, the same machine runs more locally over time.
Default models are chosen for your computer with railcall models. A 128 GB studio fits 200–250B models and keeps about 98% at home; an 8 GB laptop fits 12–15B and still keeps about 80% local. Nothing stays entirely local: the hardest answers always overflow, and that is the part you meter. Either way the overflow is metered against the budget you set, and your statement shows every request that left and why.
A laptop does not have to settle for the models it can hold. With capacity routing, the job goes to the machine in your org with headroom, runs there on hardware you own, and the receipt comes back. Everyone gets the strongest local compute in the building, for free.
Everything on your own machines is free. These are the paid services, and only these.
Everything on your own machines is $0 and never counted. The only three things with a price are seats on the hub (sophisticated workflows with multi-level, airlocked human approval), agents we host, and the overflow. Drag the sliders.
Everything local is $0 and never counted. Overflow is metered against a budget you set, priced here at the top model’s class (Fable / Astra) so it matches the competitors’ AI line; rounded to $5. 500 overflow flows a month are included.
Not a free tier. The whole runtime, on every machine you own, with no account for anything local.
Illustrative capture; receipt counts are examples, the commands are real.
Every action is a signed receipt. Every send goes through an airlock: preview, approve, execute. A company-wide switch stops everything, on every machine, and it is free at every tier.
Receipts are tamper-evident and chain-linked; anyone can verify them offline with one command. Sensitive changes need a second human and an approve code.
20 people, 10 machines, 20 always-on agents, 20,000 AI-touching steps. The same work bought from the usual stacks, each vendor's meter counted the way that vendor counts it, and every AI step at the top model's list.
7× cheaper than the Copilot stack, 13× than the Zapier stack, 22× than the n8n stack. The AI line is the point: Zapier and n8n send every one of the 20,000 AI steps to the model’s API at list (Fable 5.1 / GPT-6 Astra, $10 in / $50 out per 1M) on your own key, $1,400 a month; Copilot Studio meters them as agent actions in credits, $836. Their workflow runs are counted at each vendor’s cheapest published meter: Zapier’s Team tier billed annually, n8n’s self-hosted Business tier plus its published overage bucket, Copilot Studio’s agent flow actions. Our overflow is priced at the same top class, and only the tenth that leaves your machines is counted. None of them answers on your machines first by default; n8n self-hosted can point a node at a local model, but you build the routing, the fallback and the metering yourself.
Your machines sit idle most of the day. RailCall puts them to work and you stop paying cloud services to do what your own hardware can do.
The hardware is sunk and was about 80% idle. Run it flat out and the electricity is the whole cost. Cursor bills third-party model usage at list plus $0.25 per million tokens, inside and beyond an included pool of unpublished size, so its line is a ceiling, marked ≤; the seat products run the same 20,000 AI steps through the API at list, $1,400. Our overflow line is $210 for the same work.
A three-pane Studio, a CLI, and an MCP server so Claude Code, Cursor or any MCP client drives RailCall under the same airlock.
Build, approve and run in three panes. Everything local, nothing leaves without a receipt.
Your coding agent calls RailCall tools; writes still stop at the airlock for a human.
A governed registry of 1,000+ nodes plus any MCP tool. Bring your own local models.
Build it once; it runs on your machine forever at $0. Share it across the org on the hub.
One line on your machine. No account until you want a hub.
Install free Ask on Discord