HTTP-only client with 99% memory reduction#25
Open
mattheworiordan wants to merge 9 commits intoSawyerHood:mainfrom
Open
HTTP-only client with 99% memory reduction#25mattheworiordan wants to merge 9 commits intoSawyerHood:mainfrom
mattheworiordan wants to merge 9 commits intoSawyerHood:mainfrom
Conversation
Add serveWithExternalBrowser() that connects to an existing browser via CDP instead of launching Playwright's Chromium. Key features: - Connect to any browser with CDP enabled (Chrome for Testing, Chrome Beta, etc.) - Auto-launch browser if not running (with BROWSER_PATH env var) - Browser stays open after server stops (user manages lifecycle) - No extension required - direct CDP connection New files: - src/external-browser.ts - Core implementation - scripts/start-external-browser.ts - Startup script Use case: Local development with visible browser automation where you want to inspect results after automation completes.
When multiple AI agents run browser automation tasks in parallel, they need separate HTTP API ports while potentially sharing the same browser instance. This adds automatic port allocation to avoid conflicts. Key changes: - Add port-manager.ts for dynamic port allocation (range 9222-9300) - Server tracking via ~/.dev-browser/active-servers.json - PORT=XXXX output for agent discovery - Config file support at ~/.dev-browser/config.json - Update both standalone and external browser modes Architecture: Agent 1 → server (port 9222) ┐ Agent 2 → server (port 9224) ├→ Shared Browser (CDP 9223) Agent 3 → server (port 9226) ┘ See docs/CONCURRENCY.md for design decisions and usage examples. Addresses concerns raised in PR SawyerHood#15 about single-point congestion.
When a dev-browser server crashes, its Chrome browser may still be running on the CDP port. This adds smart cleanup to detect and terminate orphaned browsers before launching new ones. Key changes: - Enhanced ServerInfo structure to track CDP port and mode - Added detectOrphanedBrowsers() to find browsers with no registered server - Added cleanupOrphanedBrowsers() to safely terminate orphans - Standalone mode now cleans orphans on startup (before launching browser) - External mode tracks CDP port but doesn't clean (browser is intentionally external) This restores crash recovery functionality that was previously in start-server.ts, but in a smarter way that respects multi-agent scenarios.
- Add ~/.dev-browser/config.json for browser configuration - Auto-detect Chrome for Testing on macOS/Linux/Windows - Add --standalone flag to force Playwright mode - Skip npm install when dependencies unchanged (hash check) - Rename port-manager.ts to config.ts with browser config - Let browser use default profile unless userDataDir explicitly set - Simplify SKILL.md documentation with single startup flow
- Add Map-based page lookup in client.ts for O(1) targetId resolution
(eliminates 11ms CDP session scan per lookup)
- Add conditional npm install in server.sh using package-lock hash
(skips 500-2000ms when dependencies unchanged)
- Add TypeScript pre-compilation with esbuild
(500ms faster startup using node vs tsx)
- Include mode in POST /pages response to eliminate extra HTTP round-trip
- Add benchmark.ts script for measuring performance
Key improvements:
- Page lookup: 11ms → 0ms (after first access populates registry)
- Server startup: ~700ms faster (pre-compiled + conditional npm)
- HTTP requests: 1 fewer per getPage() call
Enables agents to use dev-browser without Playwright dependency by moving
page operations server-side and providing a thin HTTP client.
Server changes:
- Add HTTP endpoints for page operations in index.ts and external-browser.ts:
- POST /pages/:name/navigate - navigate to URL
- POST /pages/:name/evaluate - execute JavaScript
- GET /pages/:name/snapshot - get AI-friendly ARIA snapshot
- POST /pages/:name/select-ref - get element info by ref
- POST /pages/:name/click - click element by ref
- POST /pages/:name/fill - fill input by ref
- Add new API types (EvaluateRequest/Response, NavigateRequest/Response, etc.)
New lightweight client:
- client-lite.ts: HTTP-only client (~30KB import vs ~12MB for Playwright)
- No Playwright dependency required on agent side
- Same interface as full client for easy migration
Testing:
- http-api.test.ts: 21 tests for request validation and registry logic
- client-lite.test.ts: 24 tests for HTTP client behavior
- test-http-api.ts: Manual integration test script
- memory-benchmark.ts: Measures memory savings (60% heap reduction for 10 agents)
Memory impact (10-agent scenario):
- Before: 238 MB heap (each agent imports Playwright)
- After: 95 MB heap (server has Playwright, agents use HTTP)
- Savings: 143 MB (60% reduction)
Phase 3 of performance optimization: make client-lite the primary client path. Changes: - Extract shared HTTP routes into http-routes.ts (removes ~400 lines duplication) - Add HTTP endpoints: screenshot, set-viewport, wait-for-selector, info - Update client-lite.ts with methods for all page operations - Update SKILL.md to document client-lite API - Add deprecation notice to client.ts pointing to client-lite - Add note to scraping.md for advanced Playwright usage Benefits: - Client memory: 12.4MB → 30KB (99.8% reduction) - No Playwright dependency on client side - All page operations via HTTP to server
b503733 to
e4ebba3
Compare
- Change default port range from 9222-9300 to 19222-19300 to avoid
Chrome CDP port conflicts (9222 is Chrome's default debug port)
- Add automatic port discovery chain in client-lite:
1. DEV_BROWSER_PORT environment variable
2. tmp/port file written by server
3. Most recent server from ~/.dev-browser/active-servers.json
4. Default port 19222 as fallback
- Write port to tmp/port on server startup for client discovery
- Add 30-minute idle timeout to prevent zombie server accumulation
- Clean up stale server entries on startup
- Update SKILL.md with new configuration options and behavior
- Add start-external-browser.ts to build script
This fixes the issue where agents couldn't connect to dev-browser
because the client defaulted to port 9222 while the server was
dynamically assigned a different port.
- When browser.path ends with .app, automatically use `open -a` on macOS - Fail with helpful error instead of silent fallback to standalone mode - Add path validation for user-specified browser paths - Document .app bundle behavior in SKILL.md This ensures consistent browser behavior and proper Dock integration when using launcher apps that handle CDP flags internally.
mjdaly
added a commit
to mjdaly/dev-browser
that referenced
this pull request
Jan 16, 2026
- Switch to lightweight HTTP-only client (12MB → 30KB memory per agent) - Extract shared routes to eliminate code duplication - Add complete HTTP API for all page operations - Map-based page registry for O(1) lookup (was O(n) CDP scan) - Auto-shutdown after 30 minutes of inactivity - Stale server cleanup on startup Source: SawyerHood#25 Co-authored-by: Matthew O'Riordan <matthew@ably.com>
mjdaly
added a commit
to mjdaly/dev-browser
that referenced
this pull request
Jan 16, 2026
- Switch to lightweight HTTP-only client (12MB → 30KB memory per agent) - Extract shared routes to eliminate code duplication - Add complete HTTP API for all page operations - Map-based page registry for O(1) lookup (was O(n) CDP scan) - Auto-shutdown after 30 minutes of inactivity - Stale server cleanup on startup Source: SawyerHood#25 Co-authored-by: Matthew O'Riordan <matthew@ably.com>
4 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Whilst using dev-browser, especially with concurrent agents, I found things slow to start at times, and resource intensive. This fixes that and makes things very snappy and lightweight.
Summary
Changes
Phase 1: Startup Optimizations
Phase 2: HTTP-Only API
New server endpoints for all page operations:
POST /pages/:name/screenshot- capture screenshotsPOST /pages/:name/set-viewport- set viewport sizePOST /pages/:name/wait-for-selector- wait for elementsGET /pages/:name/info- get URL and titlePhase 3: Client-Lite
client-lite.ts- HTTP-only client, no Playwright dependencyhttp-routes.ts(~400 lines deduplication)Memory Benchmark
client-lite: 29.6 KB import overhead
Playwright: 12.4 MB import overhead
Reduction: 99.8%
Test plan
Note:
This Depends on multi-agent-concurrency #24. However, this PR is asking for merge to main, and thus includes commits from that branch. This is unavoidable because of Github :(