Case Studies & Wins

Real results from AI productivity and developer experience implementations.

better-browser-mcp: any AI agent, in your Chrome

Open source (MIT)

72% fewer tokens

Context

Browser agents spend most of their tokens looking at pages. In my own sessions they took 1,849 screenshots and read the page as text 7 times. Claude in Chrome gives Claude your browser, but it spends a lot of tokens and calls per task, and it works only for Claude.

Action

  • Built an MCP server and a Chrome extension that give any MCP client (Codex, Claude Code, others) your real Chrome profile, in its own tab group, without stealing focus.
  • The page comes back as compact text with short refs; one call runs a batch of steps and returns only what changed.
  • Benchmarked on 16 hard pages (iframes, shadow DOM, native dialogs, a prompt-injection trap) against Claude in Chrome, Playwright MCP and chrome-devtools-mcp, with success checked by the server.

Outcome

  • 72% fewer input tokens and 3.5× faster than Claude in Chrome (30.5k vs 109.9k tokens, 11 s vs 39 s per task).
  • Passed 32 of 32 runs; Claude in Chrome passed 28.
  • On real sites (Hacker News, Amazon): between a half and a quarter of the tokens.

Takeaway

When many models fail at the same thing, fix the tool, not the prompt. A text-first browser layer with batched actions and honest results makes every agent better at the web.

Read the write-up →GitHub →

MCPBrowser AutomationChrome ExtensionCodexClaude Code

Press-inquiry response optimizer

Tech Startup

14× productivity boost

Context

The comms team handled only about 10 journalist inquiries per week. Manual inbox triage and bespoke drafting caused high toil; most leads expired before anyone replied.

Action

  • Built a multi-agent LLM system that reacts to new requests every hour, scores relevance and "draftability," pulls supporting sources, and prepares a draft structure.
  • Editors use a lightweight interface (desktop or phone) to review, add feedback, and trigger the agent to rewrite into final copy; loop repeats until approved.
  • All sensitive fields stay local—nothing exposed to external LLMs. A human-in-the-loop editor remains the final gate.

Outcome

  • Response time dropped from ~7 days to <8 hours.
  • Team now sends 20+ tailored responses per day (≈14× productivity boost).
  • Best single-day record so far: 11 pitches sent.
  • Manual effort on this task cut by roughly 90 percent.

Takeaway

Decoupling a high-volume text pipeline into triage, context gathering, draft, and human review steps—then automating each with cooperating agents—turns a slow, error-prone process into a rapid, scalable one while preserving quality and compliance. The same pattern can lift productivity in any stream of repetitive knowledge work such as RFPs, support tickets, or bug-report triage.

LLMMulti-AgentCommunicationsAutomationHuman-in-Loop

Ready to see similar results?

Let's discuss how AI productivity and developer experience improvements can transform your team's workflow.

Let's Talk