Open source · Self-hosted on your own Google Cloud

Even your AI needs
a second opinion.

Your coding agent is confident. Confidently wrong, sometimes. Ask‑a‑Friend lets it phone a friend — a completely different model that reviews the work, finds what the first one missed, and reports back. All in one tool call.

Watch it work New to MCP? Start here
Setup time
~2 min
Clients supported
5+
Your code leaves scrubbed
Always
Your agent Shipping this auth middleware. Looks good to me! 
if (token === process.env.API_SECRET)
calling ask_a_friend…
The friend (different model) Hold on. That's a timing attack — === short-circuits on the first wrong byte. An attacker can guess your secret one character at a time.
Use crypto.timingSafeEqual()
…good catch. Fixing that now.

First things first

Hold on — what's an MCP?

No jargon. Here's the whole idea in about thirty seconds.

It's a USB port for AI.

Before USB, every device needed its own weird connector. MCP — the Model Context Protocol — does for AI what USB did for hardware. You build a capability once, and every MCP-compatible app can plug into it: Claude, ChatGPT, Cursor, Gemini CLI, and whatever ships next month.

  1. 1

    You ask your agent something

    "Review this before I merge it." Normal request, in whatever app you already use.

  2. 2

    It notices it has a tool for that

    Your agent sees an ask_a_friend tool is available and decides to call it — no prompting gymnastics required.

  3. 3

    The tool does real work

    Our server scrubs your secrets, routes to a different model, and streams the answer back into your chat. You never leave the app.

That's it. MCP is just the plug shape. Ask-a-Friend is what you plug in.

The actual problem

One model means one blind spot.

Asking a model to check its own work is like proofreading your own essay at 2am. It re-reads what it meant to write.

Reviewing itself

  • Same training data, same habits
  • Repeats the assumption that caused the bug
  • Agrees with itself — confidently
  • You get a rubber stamp, not a review

Phoning a friend

  • Different model family, different instincts
  • Arrives with zero context to defend
  • Disagrees when you're wrong
  • You get an actual second pair of eyes

Your secrets never leave

API keys, tokens, private keys, emails, and /Users/you/… paths are redacted before the request goes out, then restored in the reply.

It's your cloud

You deploy it to your own Google Cloud project. There's no middleman service, no shared instance, and no one else's logs.

It doesn't ask twice

Identical questions hit a SHA-256 cache instead of the model, so repeat reviews are instant and free.

Live demo

Watch a question make the round trip.

This is a faithful simulation of the real pipeline — same stages, same order. It starts on its own; pick a different question any time.

opus-5-5

You send

What happens

  1. Check who's askingVerifying token…
  2. Block prompt injectionScanning…
  3. Scrub your secretsRedacting…
  4. Asked this before?Cache check…
  5. Pick the right friendResolving…
  6. Friend thinks it overCalling model…
  7. Put your secrets backRestoring…

The friend says

Starting…
Took—
Friend—
Tokens—
Cache—

The roster

Meet the friends.

Ask for one by name, or let it route automatically based on what you're asking.

Claude Opus 5.5

opus-5-5

The thorough one

Deep reasoning for security audits and architecture critiques. Thinks hard, searches the live web, and writes up to 128K tokens of findings.

  • Adaptive thinking
  • Web search
  • 1M context

Claude Sonnet 5

sonnet-5

The quick one

Same family, faster turnaround. Good when you want a sanity check on structure rather than a full forensic audit.

  • Fast reviews
  • Web search
  • 1M context

Gemini 3.8 Flash

gemini-3.8-flash

The other perspective

An entirely different lineage, which is the whole point. Grounded in Google Search, and the automatic fallback if Claude is busy.

  • HIGH thinking
  • Search grounding
  • Auto-failover

Two minutes

Introduce your agent to its new friend.

Deploy to your own Google Cloud project, then paste one config. Pick your app:

1

Put it in your cloud

git clone https://github.com/mbettan/ask-a-friend-mcp.git
cd ask-a-friend-mcp
./deploy.sh <YOUR_GCP_PROJECT_ID>

The script enables the APIs, generates your API key, and prints your server URL.

2

Tell your agent about it

Replace your-cloud-run-url.run.app with the URL the deploy script printed.

That's the whole thing.

Next time your agent is about to ship something questionable, it'll ask a friend first.

Get it on GitHub ↗ Read llms.txt
Copied!