PROBUS
PROBUS
New Probus works inside Claude Code, Cursor and any MCP agent

Catch the Solana bugs your AI didn’t.

Security checks for Solana programs, the code that moves money around them, and the agents that sign, built for builders who ship with AI. Scan a repo free, then keep it monitored as your code changes.

No card needed.

FOR AI AGENTS

Your agent writes the code. Now it can check it too.

Connect Probus to Claude Code, Cursor or any MCP client. Your agent scans your Solana repo, reads the findings, fixes them, pushes, and re-scans to confirm the fix. Same checks, same limits as the website, and it can never spend your money.

  • Scan, read findings and get a paste-ready fix brief
  • Re-scan after a fix and see the score move
  • Watch monitored repos and write your listing
Connect your agentclaude mcp add --transport http probus https://useprobus.com/mcp
› Scan my game backend with Probus and fix what it finds
probus · start_scan dawnfall-labs/dawnfall-backend
probus · get_scan complete · score 39
probus · get_fix_brief 2 critical, 1 high
  critical  verify.ts:41  payment recipient never checked
  critical  relay.ts:12   fee-payer relay signs anything
  high      rewards.ts:9  client-reported score paid out
Fixed all three, pushed 4f2c1a9.
probus · start_scan → score 94 · Strong

How it works

1

Scan free

Connect a repo and get a score and your most serious finding in minutes.

2

Fix with your AI tool

Every finding comes with file, line and a fix brief to paste into Claude Code, Cursor or Nanocoder.

3

Keep it monitored

Probus re-checks every week, so a new commit can't quietly undo your fixes.

A generic code review doesn’t know what a signer check is.

Ask a general-purpose AI to review a Solana program and it reasons from web and Ethereum patterns, because that’s most of what it’s seen. The failure modes here are structurally different.

Missing signer check

Instructions run against whatever accounts are passed in. Without an explicit signer requirement, anyone can call it as anyone.

Unchecked account ownership

Read an account without checking your program owns it, and a lookalike can be swapped in and trusted.

Colliding PDA seeds

Change seed structure in an upgrade without versioning it, and new addresses can collide with old ones.

And in the code around the program

Payment that never happened

A verifier that checks the amount but not the recipient, mint or signature’s history pays out for a transfer that went somewhere else, or went twice.

A key that signs anything

A fee-payer relay or signing endpoint that co-signs whatever the client sends lets the client add a transfer out of your wallet.

A score the client made up

Rewards paid onchain for a score, result or balance the client reported are rewards anyone can claim.

An agent that signs what it’s told

If a prompt, a tool result or a token’s description can reach the code that signs, so can an attacker. Limits written in a prompt aren’t limits.

The rule pack is built around findings like these because they’re the ones raw prompting reliably misses, not the ones any tool would catch.

What Probus checks, and what it doesn't

Every report lists what its scan covered, so a clean score never reads as more than it is.

Onchain programs, Anchor and native

Signers, account ownership, PDAs, account confusion, re-initialisation and closing, CPIs and arithmetic.

The offchain money path

Payment verification, committed and exposed keys, signing services, transaction building, privileged routes, client-reported state, and AI agents that hold keys.

General web security

Login flows, XSS, dependency CVEs and the like. Not what Probus is for: use a general-purpose scanner alongside it.

See exactly what you'd get

We ran Probus against sealevel-attacks, a public repo built to teach the most common Solana vulnerabilities. The full report is public.

Flagged all 11 insecure programs. No findings on the fixed versions.

Read the sample report

What happens to your code

✓Probus only reads the repos you choose, and never changes your code.
✓Each scan reads your code at one commit, for that scan only. The parts it reviews go to the AI model named in our privacy policy.
✓We keep your findings and short excerpts around them, not your repository.
✓Reports are private to you. Only what you choose to share, like a badge, is public.

Our privacy policy

Who’s behind Probus

Ben Parry

Building private, local and decentralised AI infrastructure.

GitHubX

Will Lamerton

Tech entrepreneur, building Probus at Ava Technologies.

GitHub

Ava Technologies

Probus is built by Ava Technologies in the UK. The engine underneath is open source: read exactly how your code is checked.

Sentinel on GitHub

More from Ava

Engineer review

An Ava engineer checks your findings and writes a summary to share before a launch, raise or grant application.

Ask about a review

For programmes

Sponsor monitoring for every team in your hackathon, accelerator or grant cohort.

For organisers

Fixed-price fixes

Ava engineers fix Probus findings for a price agreed up front.

Get a quote

No automated tool catches everything, including this one. Probus is a middle layer between no checks and a full audit, not a replacement for one. See our terms.