Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
/**
* E2E Tests: Desktop Tools (X11 + Browser)
*
* Prerequisites:
* 1. Minikube running with `make minikube-setup`
* 2. Host CRD `chatllm` has `spec.desktop: { browser: true, x11: true }`
* 3. mcp-host pod uses `clerum/mcp-host-desktop:test` image
* 4. Port-forward active: `kubectl port-forward -n mcp-host svc/chatllm 8080:8080`
* 5. LLM provider is vision-capable (e.g., OpenAI gpt-4o)
*
* Run:
* cd tests/e2e && npx vitest run mcp-host/desktop-tools-e2e.test.ts
*/
import { beforeAll, describe, expect, it } from 'vitest'
import { execSync } from 'child_process'
import { createRequire } from 'module'
import { MCP_HOST_URL, fetchJson, healthCheck, sleep, waitForIdle } from '../helpers'
const require = createRequire(import.meta.url)
const jwt = require('jsonwebtoken')
const TASK_TIMEOUT = 120_000 // 2 min — LLM + tool execution
let authToken: string
/** Generate a JWT signed with the cluster's actual private key. */
function generateClusterJwt(): string {
// Extract the RPC JWT private key from the cluster
const keyB64 = execSync(
"kubectl get secret control-api-secrets -n control-plane -o jsonpath='{.data.CONTROL_API_RPC_JWT_PRIVATE_KEY}'",
{ encoding: 'utf-8' }
).replace(/'/g, '')
const privateKey = Buffer.from(keyB64, 'base64').toString('utf-8')
return jwt.sign(
{
sub: 'e2e-desktop-test',
typ: 'user',
teamId: 'e2e-team',
scopes: ['host:message:invoke', 'host:status:read', 'host:task:read', 'host:activity:read'],
hostRefs: ['chatllm'],
jti: `e2e-desktop-${Date.now()}`,
},
privateKey,
{
algorithm: 'RS256',
issuer: 'control-api',
audience: 'rpc-proxy',
expiresIn: '1h',
}
)
}
/** Authenticated fetch helper. */
async function authedFetch(
url: string,
opts?: RequestInit
): Promise<{ status: number; data: any }> {
const headers: Record<string, string> = {
'Content-Type': 'application/json',
Authorization: `Bearer ${authToken}`,
...((opts?.headers as Record<string, string>) ?? {}),
}
const res = await fetch(url, { ...opts, headers })
const text = await res.text()
let data: any
try {
data = JSON.parse(text)
} catch {
data = text
}
return { status: res.status, data }
}
/**
* Send a message and wait for the task to complete.
*/
async function sendAndWaitForResult(
content: string,
timeoutMs = TASK_TIMEOUT
): Promise<{ response: string; toolsCalled?: string[] }> {
const { data } = await authedFetch(`${MCP_HOST_URL}/v1/runtime/messages`, {
method: 'POST',
body: JSON.stringify({
content,
hostRef: 'chatllm',
channelId: 'e2e-desktop',
sender: 'e2e-desktop-test',
channelType: 'api',
timestamp: new Date().toISOString(),
messageId: `e2e-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`,
async: true,
}),
})
if (data.response) return data
if (data.taskId) {
const start = Date.now()
while (Date.now() - start < timeoutMs) {
const res = await authedFetch(`${MCP_HOST_URL}/v1/runtime/tasks/${data.taskId}/result`)
if (res.data?.status === 'completed' || res.data?.response) {
return res.data
}
await sleep(2000)
}
throw new Error(`Task ${data.taskId} did not complete within ${timeoutMs}ms`)
}
// Fallback: wait for idle
const start = Date.now()
while (Date.now() - start < timeoutMs) {
const res = await authedFetch(`${MCP_HOST_URL}/v1/runtime/status`)
if (res.data?.agent?.state === 'idle') return data
await sleep(1000)
}
return data
}
describe('Desktop Tools E2E', () => {
beforeAll(async () => {
// Generate auth token from cluster secrets
authToken = generateClusterJwt()
// Health check (authenticated)
const { status } = await authedFetch(`${MCP_HOST_URL}/v1/runtime/health`)
if (status !== 200) {
throw new Error(
'mcp-host is not reachable. Ensure port-forward is active: ' +
'kubectl port-forward -n mcp-host svc/chatllm 8080:8080'
)
}
// Wait for agent to be idle
const start = Date.now()
while (Date.now() - start < 30_000) {
const res = await authedFetch(`${MCP_HOST_URL}/v1/runtime/status`)
if (res.data?.agent?.state === 'idle') break
await sleep(1000)
}
}, 60_000)
describe('desktop_screenshot (X11)', () => {
it(
'takes a screenshot and the LLM describes the desktop',
async () => {
const result = await sendAndWaitForResult(
'Take a screenshot of the desktop using the desktop_screenshot tool. ' +
'Briefly describe what you see.'
)
expect(result.response).toBeTruthy()
const response = result.response.toLowerCase()
expect(
response.includes('desktop') ||
response.includes('screen') ||
response.includes('xfce') ||
response.includes('applications') ||
response.includes('icons')
).toBe(true)
},
TASK_TIMEOUT
)
})
describe('desktop_click (X11)', () => {
it(
'clicks the Applications menu and takes a screenshot',
async () => {
const result = await sendAndWaitForResult(
'Click on the Applications menu at coordinates x=50, y=8 using desktop_click, ' +
'then take a desktop_screenshot. Describe what menu items you see.'
)
expect(result.response).toBeTruthy()
const response = result.response.toLowerCase()
expect(
response.includes('terminal') ||
response.includes('file manager') ||
response.includes('web browser') ||
response.includes('settings') ||
response.includes('menu')
).toBe(true)
},
TASK_TIMEOUT
)
it(
'closes the menu by clicking elsewhere',
async () => {
await sendAndWaitForResult(
'Click at coordinates x=500, y=400 using desktop_click to close any open menus. ' +
'Just confirm the click was successful.'
)
await sleep(1000)
},
TASK_TIMEOUT
)
})
describe('browser_open (Playwright)', () => {
it(
'opens a URL and describes the page',
async () => {
const result = await sendAndWaitForResult(
'Open the browser and navigate to https://example.com using the browser_open tool. ' +
'Describe what you see on the page.'
)
expect(result.response).toBeTruthy()
const response = result.response.toLowerCase()
expect(
response.includes('example') ||
response.includes('domain') ||
response.includes('page') ||
response.includes('website')
).toBe(true)
},
TASK_TIMEOUT
)
})
describe('browser_get_content (Playwright)', () => {
it(
'extracts text content from a page',
async () => {
const result = await sendAndWaitForResult(
'Open https://example.com using browser_open, then use browser_get_content ' +
'to get the text content of the page. Tell me the exact heading text.'
)
expect(result.response).toBeTruthy()
const response = result.response.toLowerCase()
expect(response).toContain('example domain')
},
TASK_TIMEOUT
)
})
describe('combined workflow', () => {
it(
'creates an HTML file and previews it in the browser',
async () => {
const result = await sendAndWaitForResult(
'Do the following steps:\n' +
'1. Use shell_exec to create a file at /config/workspace/test.html with content: ' +
"'<html><body><h1>Hello from Clerum Agent</h1><p>Desktop tools work!</p></body></html>'\n" +
'2. Use browser_open to open file:///config/workspace/test.html\n' +
'3. Describe what the page looks like.'
)
expect(result.response).toBeTruthy()
const response = result.response.toLowerCase()
expect(
response.includes('hello') ||
response.includes('clerum') ||
response.includes('agent') ||
response.includes('desktop tools')
).toBe(true)
},
TASK_TIMEOUT
)
})
})