Coasty Review 2026: An API That Lets Agents Click Through Software With No API

A lot of the software that actually runs a business has no usable API. Payer portals in healthcare, accounting desktops, internal tools that only talk to a mouse and keyboard — the important work lives behind interfaces that were never designed to be automated. The traditional answer is RPA: record a sequence of clicks and replay it. That works right up until a button moves, a pop-up appears, or the page loads slowly, and then the whole script falls over. Coasty, a YC S26 startup from founders Nitish and Prateek, is betting that vision-based computer-use agents are the better replacement.

Coasty is an API for building computer-use agents. You send it a plain-English task, point it at an environment — a managed virtual machine or your own desktop — and its agent operates the interface exactly like a human would: looking at the screen, deciding what to do, moving the mouse, typing, checking the result, and repeating until the job is done. There is no DOM access, no accessibility tree, no per-app integration to build. If a browser, a remote desktop, or a twenty-year-old Windows app can be seen, Coasty can drive it.

What Coasty Does

At its core, Coasty replaces brittle click-replay RPA with a perception-driven agent loop. The agent captures the current screen as pixels, decides the next action, executes it through mouse and keyboard, verifies the action produced the expected state, and then loops. Because it never reads the underlying code — just the pixels — the same API can operate a web app, a remote desktop, and a legacy Windows application without any connector per system.

When a run finishes, Coasty returns a structured run record: the final status, any extracted outputs (for example a confirmation number), a replay URL, and a timestamped event log of every step. For sensitive workflows you can declare invariants — “the patient name must match the source document” — and pause the run for a named human reviewer before an irreversible step, like a claim submission. Each run executes inside an isolated virtual machine, so sessions do not share infrastructure and data does not leak between them.

Use Cases

• Healthcare prior-authorization. Teams submit prior-auth through payer portals that have no API. Coasty can open the patient record, enter authorization data, refuse to submit if the member ID or procedure code does not match, and return the confirmation number — with a human approval gate before the final push.
• Data entry across desktop apps. Accounting and operations teams still move information between internal systems, spreadsheets, and remote desktops. Coasty can bridge applications that will never expose an integration.
• Embedded computer-use in your own product. Through the REST /v1/predict endpoint and a 26-tool MCP server, developers can let their own agents or assistants trigger Coasty tasks over the standard Model Context Protocol.
• QA and form-flow testing. The agent can walk checkout flows, catch bugs, and file detailed reports without a bespoke test harness per UI.

Key Features

Vision-based computer use, no DOM needed

The agent reads the screen as pixels and acts with mouse and keyboard, so it works on legacy desktop software, internal tools, and CAPTCHA-style flows that have no API or only an incomplete one. It adapts when the UI changes, unlike selector-based RPA.

REST API plus MCP server

The primary integration is a single POST /v1/predict with a plain-English task. A 26-tool MCP server exposes browser, terminal, file, and screenshot primitives so other assistants (Claude, GPT, and similar) can call Coasty over the Model Context Protocol.

1,000+ OAuth app integrations

A Composio-style integration layer ships more than 1,000 OAuth connections — Gmail, Outlook, Salesforce, HubSpot, Linear, Jira, GitHub, Slack, Stripe, Shopify, and many more. Tasks can call them by name, using direct APIs when available and computer-use when not.

Verification layer and human-approval gates

Developers can declare invariants and pause a run for a named human reviewer before irreversible steps. The system tracks the expected workflow state and re-plans when the application diverges, which is what makes it usable in regulated environments.

Isolated VMs with a full audit trail

Every run executes in an isolated VM, and the output includes a replay URL and a timestamped event log. That gives an evidence trail of exactly what the agent did — useful when the cost of a silent mistake is high.

Pricing

Coasty’s pricing is described in two related ways. The product page lists a base of about $0.05 per /v1/predict call, with self-serve plans at roughly $19/month (Starter), $50/month (Plus), and $99/month (Unlimited). In the Launch HN thread the founders describe billing as based on agent runtime and workflow volume, with separate pricing for dedicated environments and enterprise deployments. The exact cost model for production workloads is not fully nailed down across these sources, so pilot it on a contained task and measure before committing.

Common Questions

Is Coasty open source? No. Coasty is closed-source and proprietary, with no public GitHub repository. There is no open-source license to inspect, which is a hard gate for teams that require source access or auditable code before adoption.

How does it compare to RPA like UiPath or Power Automate? Those tools are powerful but rely on recorded click-replay and per-bot licensing that can run into the tens of thousands of dollars per worker per year, and they break when the UI shifts. Coasty argues vision-based agents are more resilient and need no per-app integration — at the cost of being a younger, closed-source vendor.

Can I trust the benchmark? Coasty reports 85.60% completion across 369 real-world OSWorld tasks, which it describes as the top score at launch; an independent source cites roughly 82.81% on the official leaderboard. The benchmark is self-reported, so validate reliability on your own applications before depending on it.

Verdict

Coasty solves a real problem that keeps getting sharper: a huge amount of enterprise software has no usable API, and traditional RPA breaks the moment a button moves. Its vision-based, verification-gated computer-use agent is a credible successor to click-replay RPA, and the REST plus MCP surface makes it genuinely embeddable. The honest caveats are maturity, transparency, and cost clarity. It is closed-source with no public repo, so there is no license to audit — an enterprise-adoption hard gate. The flagship benchmark is self-reported, it is a YC S26 startup with only modest launch traction so far, and the pricing story mixes per-call and runtime/volume models. For the 9bests audience this is a solid, try-it-on-your-own-apps 7.0/10: promising infrastructure worth piloting inside a contained workflow, but not yet proven at scale or open enough to drop into regulated production blindly.