The most observable
Android harness
an agent can use.

50+ tools. Network interception before encryption. App logs. Performance metrics. Cloud-hosted. One npx command.

The Problem

Your agent sees the screen.
It misses what's underneath.

Other harnesses tap and screenshot. Doksi reads network responses, application logs, and app state. It catches the bugs the screen doesn't show.

Other harness (mobile-mcp, agent-device etc.)

Test Passed

Doksi Harness

Bug caught
How it works

One command.
The agent does the rest.

Add Doksi to your MCP config. From that point, your agent boots devices, installs builds, navigates, and reads every log on its own.

Works with Claude Code, Cursor, Codex, Windsurf, Copilot and others.
From $0.04/min. No local setup. No subscription.

what you do (once)

what the agent does

Capabilities

50+ tools. Four categories.
Full device control.

Observation, interaction, app lifecycle, and environment simulation. Network logs with full request and response bodies. Application logs with crash detection. Performance and memory profiling. SMS injection, phone call simulation, GPS spoofing, network throttling, fingerprint auth, dark mode, font scale, battery simulation. Everything your agent needs to operate a real Android device.

Benchmark

Benchmarked against popular
open source Android harness.

50 runs across 3 harnesses. Same app, same 11 test cases, 2 bugs planted. One harness found them every time.

50
Total runs
100%
Device MCP bug detection
17%
mobile-mcp bug detection
2%
agent-device detection
Doksi MCP
mobile-mcp
agent-device
0100200300400500600700800850
Average time (seconds)
Doksi MCPDoksi MCP (provision time included)mobile-mcpagent-device

Tested using Claude Opus 4.6 on Doksi TestBench app, 11 test cases per run, 2 bugs planted. Doksi runs include ~90s cloud device provisioning. Competitors ran on a local emulator with zero provisioning time.

Works with your agent

One config. Every AI agent.

Add Doksi to any MCP-compatible agent in under a minutes. Your API key, one npx command, done.

Cursor
cursor --add-mcp '{"name":"doksi-device",
  "command":"npx",
  "args":["-y","@doksi/mcp-device"],
  "env":{"DOKSI_API_KEY":"your-key"}}'
Claude Code
claude mcp add doksi-device \
  --env DOKSI_API_KEY=your-key \
  -- npx -y @doksi/mcp-device
VS Code Copilot
code --add-mcp '{"name":"doksi-device",
  "command":"npx",
  "args":["-y","@doksi/mcp-device"],
  "env":{"DOKSI_API_KEY":"your-key"}}'
Windsurf / Cline / Continue / Zed
{
  "mcpServers": {
    "doksi-device": {
      "command": "npx",
      "args": ["-y", "@doksi/mcp-device"],
      "env": { "DOKSI_API_KEY": "your-key" }
    }
  }
}
Pricing

Pay for what you use.

From $0.04 per minute. No subscription required.

$10

One-time

200 minutes · 1 concurrent device

Full device state API (50+ tools)
Network log interception
Performance + memory monitoring
Local dev tunnel (Tailscale)
Works with Claude Code, Cursor, Codex, Copilot, Windsurf
RECOMMENDED
$20

One-time

500 minutes · 2 concurrent devices

Everything in Starter
2 devices running in parallel
Faster feedback on multi-flow coverage

What does a session actually cost?

Quick smoke test (login + home screen)3 min$0.15
Full checkout flow + network assertion8 min$0.40
Edge case suite (biometric, call interrupt, 2G)18 min$0.90
Full regression across 5 flows35 min$1.75

Billing starts when the session is ready, not when the agent initiates a session.

15-minute inactivity timeout — session ends automatically so you’re never charged for idle time.

Your agent can now
close the mobile loop.

Write code. Install it. See it. Fix it. All without picking up a phone.

No local Android setup. No ADB. No device farm account.