Coasty - YC S26 AI Agent with Computer-Use API

TL;DR

TL;DR: Coasty is a cloud-hosted AI computer-use agent with a REST API ($0.05/predict), 1,000+ OAuth app integrations via MCP, and an 85.60% OSWorld benchmark score — positioned as the API-first replacement for legacy RPA tools.

Source and Accuracy Notes

This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: coasty.ai
• Launch HN thread: news.ycombinator.com/item?id=48922706
• YC batch: ycombinator.com/companies/coasty
• License: proprietary SaaS (verified on project site)

What Is Coasty?

Coasty is an AI-native computer-use agent that drives a remote, cloud-managed desktop by reading screen pixels and executing mouse/keyboard actions — the same way a human operator would. Rather than requiring per-process RPA scripts or certified-developer builds, Coasty uses AI vision to adapt when UIs change.

It ships as a REST API (/v1/predict) plus an MCP server with 26 built-in tools, making it directly embeddable into developer workflows. The company positions it as the AI-native replacement for Blue Prism, UiPath, Automation Anywhere, and Power Automate.

Key claims from the product page:

• $0.05 /v1/predict base price, with published surcharges
• 85.60% on OSWorld benchmark — ranked #1 at time of launch, across 369 real-world computer tasks
• 1,000+ OAuth-native app integrations (Gmail, Slack, Notion, GitHub, Salesforce, HubSpot, Linear, Jira, Stripe, Shopify, and 990+ more)
• Free sandbox keys
• Public OpenAPI 3.1 spec + llms.txt for LLM ingestion

How the Agent Loop Works

Coasty runs a four-phase loop on every step:

1. Observe — captures the current screen state as pixels
2. Act — decides and executes a mouse or keyboard action
3. Verify — confirms the action produced the expected result
4. Recover — backtracks and retries on failure

This is why it scores on OSWorld: the benchmark tests 369 real-world tasks (not synthetic UI DOMs), requiring genuine visual grounding to handle the variety of real applications.

The /v1/predict API

The primary integration point is a single REST endpoint:

POST https://coasty.ai/v1/predict
Authorization: Bearer <your-api-key>
Content-Type: application/json

{
  "task": "Go to GitHub and star the repository owner/repo",
  "context": {
    "url": "https://github.com",
    "goal": "Star the repo using your GitHub account"
  }
}

Pricing is $0.05 per predict call, with published surcharges (no surprise fees). Free sandbox keys are available for testing. The full OpenAPI 3.1 spec is at https://coasty.ai/openapi.json.

MCP Server: 26 Built-in Tools

Coasty ships an MCP server with 26 pre-built tools covering common automation primitives:

computer_use   # Core screen-read + action-execute tool
mcp__files__read
mcp__files__write
mcp__browser__navigate
mcp__browser__click
mcp__browser__type
mcp__slack__send_message
mcp__github__create_issue
... (26 tools total)

Real-World Task Examples

Coasty’s demos show it handling end-to-end workflows autonomously:

• Autonomous Reddit marketing — researches competitors, identifies subreddits, crafts posts, engages with comments
• Sales prospecting — finds prospective customers, researches companies, writes personalized outreach, sends emails
• QA testing — navigates checkout flows, catches bugs, files detailed reports
• Job applications — finds matching roles, tailors resumes, submits across platforms
• YC application filling — completed an entire YC S26 application autonomously

Healthcare and Enterprise

For regulated environments, Coasty offers:

• PHI stays under your control — can deploy into your own VPC; screenshots and logs stored inside your perimeter
• BAA support — available when screen content may contain PHI
• On-premises / BYOK — runs inference on your own model keys under your own provider agreement
• HITL approvals — gated steps stop and wait for a named human reviewer before claim submissions, appeal filings, auth requests, or adjustments go to a payer
• Audit log — one screenshot per action, one approval record per gated step, replayable run records

Setup

No software to download, no servers to run:

1. Sign up at coasty.ai — free sandbox keys available
2. Bring your own model API keys (BYOK model — you use your own inference provider under your own agreement)
3. POST to /v1/predict or connect via the MCP server

# First API call
curl -X POST https://coasty.ai/v1/predict \
  -H "Authorization: Bearer $COASTY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "task": "Search for runany on GitHub",
    "context": { "url": "https://github.com" }
  }'

Benchmark

Coasty claims 85.60% on OSWorld, which the product page describes as completion rate across 369 real-world computer tasks. OSWorld is a benchmark that tests AI agents on actual desktop environments (Ubuntu, macOS, Windows) using real applications — not synthetic DOM structures. This is a harder test than typical web-agent benchmarks.

OSWorld benchmark source: os-world.github.io (open research benchmark, not Coasty proprietary)

FAQ

Q: Is Coasty cloud-only or can it run on-premise? A: Cloud-hosted by default. On-premises VPC deployment is available for enterprise customers with specific compliance requirements.

Q: What model does Coasty use? A: Coasty is model-agnostic at the API layer. Enterprises can BYOK — using their own model provider under their own agreement. The inference runs on keys you own.

Q: Does Coasty handle MFA and login flows? A: Yes. The agent handles MFA through your identity controls and connects natively where a real interface exists.

Q: How does Coasty compare to browser-use or Skyvern? A: Coasty’s differentiator is its API-first design ($0.05/predict, OpenAPI 3.1, MCP server), OSWorld benchmark ranking, and 1,000+ OAuth integrations. Browser-use and Skyvern are open-source projects with different deployment models.

Q: What does the free sandbox include? A: Free keys with rate limits for testing the /v1/predict API and MCP server locally. Production pricing is pay-per-call with no monthly minimums.

Q: Does Coasty have human-in-the-loop controls? A: Yes. Gated steps (claim submissions, auth requests, appeals, adjustments) pause for a named human reviewer before anything reaches a payer. Read-only checks like eligibility and status run automatically with full logging.

Conclusion

Coasty differentiates from the broader browser-automation agent space through three concrete claims: an API-first product ($0.05/predict, OpenAPI 3.1, MCP server), a benchmark-validated computer-use engine (85.60% OSWorld), and a 1,000+ OAuth integration layer for workplace apps. The healthcare-ready feature set — on-prem deployment, BYOK, HITL gates, screenshot audit logs — is targeted squarely at the revenue cycle management market where legacy RPA has deep roots but high per-bot costs.

If you need to embed autonomous desktop automation into a product, Coasty’s API model is worth evaluating against the build-it-yourself path of combining an OSWorld-competent model with platform automation tooling.