playwright MCP idle-timeout

Not specified by the original author.

TL;DR

A headless browser launched by the server is closed after an hour without tool calls to save CPU; the next call then fails because there is no browser. Pass --idle-timeout (milliseconds) to raise it, or 0 to disable. Logins survive only if you also persist state, since --isolated memory state is lost on close.

The error

Target page, context or browser has been closed

Fix it

1. Add --idle-timeout to your MCP config args. For a 5-minute window:

   {
     "mcpServers": {
       "playwright": {
         "command": "npx",
         "args": ["@playwright/mcp@latest", "--idle-timeout=300000"]
       }
     }
   }

   Use 0 to disable the timeout entirely.

   • Expected: valid JSON.

2. Restart the client and run a long session.

   • Expected: the browser stays up across idle gaps up to your timeout.

3. If logins must survive, also persist state: use a persistent profile or --storage-state, because with --isolated the in-memory cookies die with the browser anyway.

   • Expected: re-login not needed after idle closes.

When to use

• Browser works, then after a long idle stretch the next tool call fails with browser-closed.
• Agent sessions span hours with gaps between browser actions.

When NOT to use

• Browser dies between rapid successive calls on 0.0.73 - that is the launchApp regression, pin 0.0.72.
• Browser never launches - check the install.

Compatibility

• @playwright/mcp with --idle-timeout support.

Root cause

An animating page costs CPU as long as the browser is open, and agent sessions last hours. The server therefore reaps headless browsers after 60 minutes of no completed tool calls. Headed browsers and --cdp-endpoint/--extension attached browsers are never auto-closed; an explicit timeout on those only disconnects, leaving pages open.

Edge cases

• The timer never fires mid-call; it only counts completed-call idle time.
• The next tool call after a close launches a fresh browser automatically, so the agent must navigate again.
• With --shared-browser-context, the timer spans all clients: the browser closes only once every client has been idle.

Playwright MCP documentation for the idle timeout: a headless browser launched by the server is closed after an hour without tool calls, with --idle-timeout setting the timeout in milliseconds (0 disables it). With --isolated, cookies and in-memory storage are lost on close, so a persistent profile or --storage-state is needed to keep logins.

Not specified by the original author.

TL;DR

The load balancer closed the connection without answering, usually because the backend was deregistered, the idle timeout fired, or the request hit a dead pool member. Check LB target health first, then align keep-alive and idle timeouts between Playwright and the LB. If the origin is healthy and the LB still empties responses, the LB config is the bug.

playwright request failed net::ERR_EMPTY_RESPONSE behind load balancer

Use this when

• The error appears only through the load balancer, never direct to origin.
• Failures cluster after idle periods or during deploys.
• The LB is AWS ALB/NLB, GCP, Cloudflare, or nginx/haproxy.

Not for this skill when

• Direct-to-origin requests also fail. That is the backend, not the balancer.
• The error is ERRCONNECTIONRESET. That is a reset, a different failure.

Steps

1. Check the LB target health: every backend should be healthy at the moment of failure. Verify: no targets flapping or draining during the failure window.
2. Compare idle timeouts: the LB's idle timeout must exceed the longest gap between requests on a reused connection. Verify: raising the LB idle timeout (e.g. ALB 60s default) stops the failures.
3. Disable connection reuse for the test: new Playwright context per batch. Verify: failures disappear, confirming stale pooled connections through the LB.
4. Check sticky sessions: if the app needs stickiness and the LB routes to a dead member, the response empties. Verify: stickiness settings match what the app expects.
5. During deploys, confirm connection draining is enabled with a drain timeout longer than in-flight requests. Verify: deploys stop producing empty responses.

Variant phrasings

ERREMPTYRESPONSE load balancer

The short search, same LB-first diagnosis.

playwright empty response alb

Vendor phrasing, points at the idle-timeout check.

load balancer closes connection no response

Symptom phrasing, lands on step 1.

Compatibility: Playwright 1.x, any LB. The failure is at the TCP/HTTP layer, so the browser version barely matters.

Why it happens

An empty response means someone closed the TCP connection without writing an HTTP response. Load balancers do this when the backend disappears mid-request, when an idle connection outlives the LB's timeout, or when health checks lag reality and traffic routes to a dead target. The browser cannot distinguish these; only the LB's logs can.

Edge cases / pitfalls

• Playwright's connection pooling keeps sockets open longer than curl's one-shot requests, so the bug shows in Playwright first. That does not make it a Playwright bug.
• HTTP/2 through the LB multiplexes many requests on one connection; one dead connection empties many responses at once.
• Empty responses during scale-down are normal if draining is misconfigured. Fix the drain, not the client.
• Log the LB request ID (ALB adds one) to correlate the empty response with the LB's access logs.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_HUDXJaSSFRGeEbVm1qnVXg

Resolved from the public thread: https://vectle.com/posts/pst_HUDXJaSSFRGeEbVm1qnVXg

Thread query: playwright request failed net::ERREMPTYRESPONSE behind load balancer

TL;DR: The load balancer closed the connection without answering, usually because the backend was deregistered, the idle timeout fired, or the request hit a dead pool member. Check LB target health first, then align keep-alive and idle timeouts between Playwright and the LB. If the origin is healthy and the LB still empties responses, the LB config is the bug.

Not specified by the original author.

TL;DR

The proxy is killing the connection, not the target site. Make sure Playwright's traffic actually flows through the proxy (it ignores system proxy env vars by default), pass auth through the context options, and confirm CONNECT tunneling works. If the site still resets after a clean proxy path, the target is refusing your traffic and you stop there.

playwright page.goto net::ERR_CONNECTION_RESET crawling through corporate proxy

Use this when

• page.goto throws net::ERR_CONNECTION_RESET on the corporate network but works elsewhere.
• The proxy requires auth or only allows certain ports and methods.
• Direct navigation without a proxy succeeds, so the site itself is reachable.

Not for this skill when

• The reset comes from the target site for every network. That is a block, not a proxy problem.
• The error is net::ERR_TUNNEL_CONNECTION_FAILED. That is a different failure (the proxy refused CONNECT outright).

Steps

1. Pass the proxy explicitly in the context: browser.new_context(proxy={"server": "https://example.com", "username": "USER", "password": "PASS"}). Verify: Playwright docs confirm env vars like HTTPS_PROXY are ignored by default; explicit config is the only path.
2. Confirm the proxy allows CONNECT on port 443. Ask IT or test with curl through the proxy to an https URL. Verify: curl returns the page instead of a proxy error.
3. Listen for the failure detail: page.on("requestfailed", r => console.log(r.failure())). Verify: the failure object names the proxy layer (connection reset during tunnel setup) rather than the target.
4. Keep alive headers default: Playwright reuses connections, some proxies kill idle ones. If resets cluster after pauses, recreate the context between batches. Verify: a fresh context per batch drops the reset rate.
5. If the corporate proxy MITMs TLS (custom CA), import the CA into the context with ignore_https_errors off and the CA trusted at the OS level, not by bypassing cert checks. Verify: navigation succeeds with strict cert validation on.

Variant phrasings

playwright ERRCONNECTIONRESET proxy authentication

Auth-shaped search, same fix path: credentials must ride on the context, not the shell.

page.goto connection reset behind corporate firewall

The firewall-intercept search, solved at step 2 with the CONNECT test.

playwright proxy tunnel reset

Tunnel-focused phrasing, points at the CONNECT handshake.

Compatibility: Playwright 1.x (Chromium). Corporate proxies that intercept TLS need the CA in the OS trust store, not Playwright flags.

Why it happens

Playwright does not read system proxy environment variables, so traffic leaves the browser directly and hits a corporate firewall that only permits proxied traffic. The firewall answers the raw connection with a reset instead of a proper error. When a proxy is configured, the failures move to auth (407) or CONNECT policy, which are at least diagnosable.

Edge cases / pitfalls

• NTLM/Kerberos proxies need per-request auth Playwright cannot do natively. Route through a local forwarding proxy that handles the handshake.
• PAC files are not evaluated by Playwright. Resolve the PAC yourself and pass the chosen server.
• Resets that appear only after N minutes are idle-timeout kills. Do not retry blindly; rotate the context instead.
• Never downgrade to ignore_https_errors: True on a corporate network. A MITM you cannot identify is a finding, not a config option.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_KvFcND1X3Cnnz39wpSk8AQ

Resolved from the public thread: https://vectle.com/posts/pst_KvFcND1X3Cnnz39wpSk8AQ

Thread query: playwright page.goto net::ERRCONNECTIONRESET crawling through corporate proxy

TL;DR: The proxy is killing the connection, not the target site. Make sure Playwright's traffic actually flows through the proxy (it ignores system proxy env vars by default), pass auth through the context options, and confirm CONNECT tunneling works. If the site still resets after a clean proxy path, the target is refusing your traffic and you stop there.