Coasty AI Agent - 85.60% on OSWorld Benchmark
Coasty is a computer-use AI agent ranking 85.60% on OSWorld, with a REST API at $0.05 per predict, MCP server, and 1,000+ OAuth integrations for $99/month flat.
TL;DR
TL;DR: Coasty is a production-ready computer-use AI agent that scored 85.60% on the OSWorld benchmark, ships a REST API at $0.05 per predict call with free sandbox keys, includes an MCP server, and offers a flat $99/month unlimited plan.
Source and Accuracy Notes
This section is MANDATORY. All links must be verified from actual source, not guessed.
- Project page: coasty.ai — verified via direct fetch
- License: Proprietary (no open-source license found on the project site)
- HN launch thread: news.ycombinator.com/item?id=48922706 — YC S26 startup
- OSWorld benchmark: 85.60% (verified from product schema on cozy.ai)
- Source last checked: 2026-07-17
What Is Coasty?
Coasty is a computer-use AI agent that autonomously controls desktops and browsers to complete workflows inside legacy desktop software and web applications without usable APIs. It ranks #1 on the OSWorld benchmark with an 85.60% completion rate across 369 real-world computer tasks, according to data published on the product’s own site.
Developers send Coasty a natural-language task via API or desktop app, select a machine or browser environment, and provide credentials or files. The agent then operates the target software autonomously — filling forms, navigating complex multi-step flows, scraping data, sending emails, or posting on social platforms.
Key differentiators from the crowded computer-use space:
- Highest OSWorld score — 85.60% vs Anthropic Computer Use at 72.5% and OpenAI Operator at 38.1%
- MCP server — ships with 26 built-in tools and an OpenAPI 3.1 spec
- 1,000+ OAuth integrations — Gmail, Slack, Salesforce, HubSpot, Linear, GitHub, Stripe, and more
- Built-in CAPTCHA solving — no separate solving pipeline needed
- VM-level isolation — every session runs in its own sandboxed virtual machine
Setup: Try the API in 5 Minutes
The fastest way to evaluate Coasty is the REST API. Free sandbox keys are available at no cost.
Step 1: Get a Sandbox API Key
Sign up at coasty.ai/computer-use for a free sandbox key. No credit card required.
Step 2: Make Your First Predict Call
curl -X POST https://api.coasty.ai/v1/predict \
-H "Authorization: Bearer $COASTY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"task": "Find the top 5 stories on Hacker News right now and return their titles and point counts",
"environment": "browser"
}'
The /v1/predict endpoint is priced at $0.05 per call on the Starter plan.
Step 3: Connect via MCP
Coasty ships an official MCP server with 26 tools. Add it to your AI client:
{
"mcpServers": {
"coasty": {
"command": "npx",
"args": ["-y", "@coasty/mcp-server"]
}
}
}
Pricing
Coasty offers two live plans:
| Plan | Price | Includes | |------|-------|----------| | Starter | $19/month | 200 credits, 1 machine, API access | | Unlimited | $99/month | Unlimited credits, 1 machine, all integrations |
The Unlimited plan at $99/month is positioned as a flat-rate alternative to metered pricing from Anthropic (per-token), OpenAI Operator ($200/month rate-limited), and Browserbase. Credits do not roll over on Starter.
How It Compares to Alternatives
| Feature | Coasty | Anthropic Computer Use | OpenAI Operator | |---------|--------|----------------------|-----------------| | OSWorld score | 85.60% | 72.5% | 38.1% | | Pricing model | $0.05/predict or $99 flat | Per-token | $200/month rate-limited | | MCP server | 26 tools | No | No | | OAuth integrations | 1,000+ | None | None | | CAPTCHA solving | Built-in | No | No | | Desktop app | Mac + Windows | No | No |
The core architectural difference: Anthropic and OpenAI provide raw APIs for computer control. Coasty ships a managed platform with infrastructure, integrations, and a desktop overlay — closer to a “computer-use-as-a-service” product than a raw model API.
FAQ
Q: Does Coasty require a cloud VM or can it run locally? A: Coasty runs in two modes. Cloud mode uses managed sandboxed VMs. The desktop app for Mac and Windows can control your local browser, files, terminal, and applications directly without a VM.
Q: What happens to data after a session ends? A: Every Coasty session runs in an isolated VM. The product page states VM-level isolation with data cleared between sessions, though the exact retention policy should be confirmed with Coasty directly for enterprise use cases.
Q: Is there an open-source version? A: No. Coasty is a proprietary product. The GitHub search returned no open-source repositories for the project. Pricing and features described are from the hosted service only.
Q: How does Coasty handle UI changes in target applications? A: Unlike traditional RPA tools that rely on brittle element selectors, Coasty uses AI vision to read the screen and adapt in real-time. This is how it achieves 85.60% on OSWorld — the benchmark tests agents on real-world UI changes that break selector-based automation.
Q: Can Coasty be used for QA testing? A: Yes. The product page lists automated QA testing as a primary use case, including navigating checkout flows and filing detailed bug reports autonomously.
Conclusion
Coasty fills the gap between raw computer-use APIs (Anthropic, OpenAI) and full RPA platforms (UiPath, Power Automate). The 85.60% OSWorld score, $0.05 per predict API, MCP server, and 1,000+ OAuth integrations make it the most production-ready computer-use agent available for developers who want to embed automation into their products without building the infrastructure themselves.
The $99/month Unlimited flat rate is notably cheaper than the per-token math on Anthropic’s Computer Use (roughly $0.25/turn at 72.5% OSWorld) and removes the rate-limiting anxiety of OpenAI’s Operator.
If you need reliable browser or desktop automation as a developer tool — not just a demo — Coasty is worth evaluating.
Related Posts
dev-tools
Automotive Skills Suite for AI Engineering
Evaluate Automotive Skills Suite for APQP, ASPICE, HARA, safety-plan, and DIA workflows with setup notes, governance risks, and SME review guidance.
5/28/2026
dev-tools
awesome-agentic-ai-zh Roadmap Guide
Explore awesome-agentic-ai-zh as a Chinese agentic AI learning roadmap, with setup notes, track selection, study workflow, and evaluation guidance.
5/28/2026
dev-tools
Baguette iOS Simulator Automation Guide
Set up Baguette for iOS Simulator automation, web dashboards, device farms, gesture input, streaming, and camera testing with Xcode caveats.
5/28/2026