SolAudit AI is a collection of open LLMs specialized in Solidity security auditing. They run on your own machine via Ollama, so your code never leaves it, and they return reentrancy, access control and oracle manipulation findings as a structured JSON report.
❯ curl -s -X POST https://solauditai.dev/api/audit -H "Content-Type: text/plain" --data-binary @Vault.sol | jq .report
⠿ solaudit-coder:7b analyzing… 212 lines · 2.8s
CRITICAL Reentrancy in withdraw() SWC-107 · L16
HIGH tx.origin used for authorization SWC-115 · L22
LOW Missing events for state changes L10, L18
risk_score: 94 / 100
Models
Every model ships as an Ollama Modelfile that layers a Solidity audit system prompt and tuned parameters on top of a proven open-source code LLM.
Balanced speed and accuracy
The default model for everyday PR reviews and CI pipelines. Its code-focused base gives it a strong grasp of Solidity syntax.
ollama create solaudit-coder:7bStep-by-step deep analysis
Reasons through call flows and state changes step by step to uncover compound bugs like reentrancy and price manipulation.
ollama create solaudit-deep:14bFinal check before mainnet
The flagship model for large protocols and multi-contract codebases, with the lowest false-positive rate.
ollama create solaudit-pro:32bLight enough for a laptop
A lightweight model for quick first-pass scans and learning. Great for on-save checks in your editor.
ollama create solaudit-lite:3b* Accuracy and speed are indicative values for comparing models. Real-world performance depends on your hardware and codebase.
How it works
No cloud API keys, no usage billing. Terminal → Ollama → security report — that's the whole pipeline.
Send your .sol file as-is with curl. Works with the local Ollama API (:11434) or this site's /api/audit proxy.
curl --data-binary @Vault.solOllama runs the audit model on your own GPU/CPU while the system prompt walks the code through an SWC Registry checklist.
ollama · temperature 0.1Get structured JSON with severity, location, SWC ID and a fix for every finding — ready for jq, CI and dashboards.
format: "json"
Why local AI
SolAudit AI is your first line of defense, catching common mistakes before a professional audit. Get instant feedback right inside your dev loop.
Analyze unreleased protocol code with confidence. All inference happens in your local Ollama runtime.
No per-token billing, so run audits on every commit and every file at zero cost.
Ollama's JSON mode always returns reports in the same schema. Automate without parsing headaches.
With just curl and jq, fail GitHub Actions or GitLab CI builds whenever a critical issue is found.
Quick start · curl
Pick a model and OS and the commands below update automatically. Just copy and paste them in order.
Install Ollama, the local LLM runtime. Once installed, the server runs in the background on port :11434.
curl -fsSL https://ollama.com/install.sh | sh
# Verify the installation (if the server isn't running: ollama serve)
ollama --version
curl http://localhost:11434/api/versionsolaudit-coder:7b is built on qwen2.5-coder:7b (4.7 GB).
ollama pull qwen2.5-coder:7bDownload the Modelfile (audit system prompt + parameters) with curl and register it as an Ollama model.
curl -fsSL https://solauditai.dev/api/modelfile/solaudit-coder-7b -o solaudit-coder-7b.Modelfile
ollama create solaudit-coder:7b -f solaudit-coder-7b.Modelfile
ollama list | grep solauditSend a code snippet straight to Ollama's /api/generate. format: "json" guarantees a structured report.
curl http://localhost:11434/api/generate -d '{
"model": "solaudit-coder:7b",
"prompt": "contract A { function kill() public { selfdestruct(payable(msg.sender)); } }",
"format": "json",
"stream": false
}' | jq -r '.response | fromjson'jq -Rs safely wraps the file contents in a JSON string and pipes it to /api/chat.
jq -Rs '{
model: "solaudit-coder:7b",
stream: false,
format: "json",
messages: [{ role: "user", content: . }]
}' Vault.sol \
| curl -s http://localhost:11434/api/chat -d @- \
| jq -r '.message.content | fromjson'Run this site with npm run dev and /api/audit calls Ollama for you. Send your .sol file as-is — no JSON escaping needed.
curl -s -X POST "https://solauditai.dev/api/audit?model=solaudit-coder:7b" \
-H "Content-Type: text/plain" \
--data-binary @Vault.sol | jq .
# Real-time streaming (NDJSON)
curl -N -X POST "https://solauditai.dev/api/audit?model=solaudit-coder:7b&stream=true" \
-H "Content-Type: text/plain" \
--data-binary @Vault.solLive example
The result of auditing a Vault contract that hides a classic reentrancy bug and tx.origin authentication with solaudit-coder:7b.
Input · Vault.sol
1// SPDX-License-Identifier: MIT2pragma solidity ^0.8.20;34contract Vault {5 mapping(address => uint256) public balances;6 address public owner;78 constructor() { owner = msg.sender; }910 function deposit() external payable {11 balances[msg.sender] += msg.value;12 }1314 function withdraw() external {15 uint256 amount = balances[msg.sender];16 (bool ok, ) = msg.sender.call{value: amount}("");17 require(ok, "transfer failed");18 balances[msg.sender] = 0;19 }2021 function sweep(address to) external {22 require(tx.origin == owner, "not owner");23 payable(to).transfer(address(this).balance);24 }25}Output · POST /api/audit
{
"model": "solaudit-coder:7b",
"duration_ms": 2814,
"report": {
"summary": "Vault is exposed to reentrancy and phishing-based owner takeover. Funds can be fully drained.",
"risk_score": 94,
"findings": [
{
"id": "SA-001",
"title": "Reentrancy in withdraw()",
"severity": "critical",
"swc": "SWC-107",
"location": "withdraw() L16-18",
"description": "External call is made before the balance is zeroed, allowing a malicious receiver to re-enter and withdraw repeatedly.",
"recommendation": "Apply Checks-Effects-Interactions: set balances[msg.sender] = 0 before the call, or use ReentrancyGuard."
},
{
"id": "SA-002",
"title": "tx.origin used for authorization",
"severity": "high",
"swc": "SWC-115",
"location": "sweep() L22",
"description": "A contract called by the owner can invoke sweep() and pass the tx.origin check.",
"recommendation": "Replace tx.origin with msg.sender and consider OpenZeppelin Ownable."
},
{
"id": "SA-003",
"title": "Missing events for state changes",
"severity": "low",
"swc": null,
"location": "deposit() L10, withdraw() L18",
"description": "Deposits and withdrawals emit no events, hindering off-chain monitoring.",
"recommendation": "Emit Deposit and Withdraw events."
}
],
"gas_optimizations": ["Declare owner as immutable", "Use custom errors instead of revert strings"]
}
}Coverage
From classic bugs to DeFi-specific attack vectors and gas optimization tips — all checked in a single request.
Repeated withdrawals caused by updating state after external calls
Missing onlyOwner, unprotected selfdestruct
delegatecall into untrusted callees
Spot price reliance, flash-loan price manipulation
Privilege takeover through phishing contracts
Ignored low-level call failures
Arithmetic in unchecked blocks or pre-0.8 compilers
Signature reuse without a nonce or chainId
Storage layout collisions in upgradeable proxies
Transaction ordering dependence, missing slippage limits
Randomness derived from block.timestamp / blockhash
Unbounded loops, halts caused by reverting external calls
API reference
Call Ollama (http://localhost:11434) directly, or use the SolAudit proxy (https://solauditai.dev) for an even simpler request.
/api/auditTakes Solidity source, audits it with Ollama and returns a JSON reportSolAudit/api/modelsAvailable models and Modelfile download URLsSolAudit/api/modelfile/:slugModelfile for ollama create (text/plain)SolAudit/api/chatChat request — pass the contract in the messages arrayOllama/api/generateSingle-prompt request — for testing short snippetsOllama| Name | Type | In | Description |
|---|---|---|---|
| code | string | JSON body | Solidity source (the entire body for text/plain requests) |
| model | string | body · query | Model to use. Defaults to solaudit-coder:7b |
| stream | boolean | body · query | If true, Ollama's NDJSON stream is passed through as-is |
Request with a JSON body
curl -s https://solauditai.dev/api/audit \
-H "Content-Type: application/json" \
-d '{
"model": "solaudit-pro:32b",
"code": "pragma solidity ^0.8.20; contract T { function f() external { selfdestruct(payable(msg.sender)); } }"
}'CI pipeline gate
# Fail CI (exit 1) if any critical / high issue is found
for f in contracts/*.sol; do
curl -s -X POST "https://solauditai.dev/api/audit" \
-H "Content-Type: text/plain" --data-binary @"$f" \
| jq -e '[.report.findings[] | select(.severity=="critical" or .severity=="high")] | length == 0' \
|| { echo "❌ $f"; exit 1; }
doneFAQ
No. SolAudit AI is a first-pass tool for catching common mistakes during development. LLMs can produce false positives and miss issues, so any contract holding real funds should also go through static and dynamic analysis such as Slither and Foundry fuzzing, plus a review by a professional audit firm.
When you call Ollama (localhost:11434) directly, your code never leaves your machine. When using the /api/audit proxy, it only travels between this Next.js server and the Ollama server configured in OLLAMA_HOST.
Yes. Ollama works on CPU alone, just more slowly — without a GPU we recommend solaudit-lite:3b or solaudit-coder:7b. Apple Silicon Macs get Metal acceleration automatically.
Yes. Add a line such as "Write description and recommendation in Korean" to your request, or append an output-language rule to the SYSTEM prompt in the Modelfile and re-run ollama create.
Sending one file at a time, within the model's context window (16K–32K tokens), gives the most accurate results. Loop over contracts/*.sol like the CI script in the API section, or send the output of forge flatten to solaudit-pro:32b.
Set OLLAMA_HOST=http://<server IP>:11434 in .env.local and start the server with OLLAMA_HOST=0.0.0.0 ollama serve. Always put a reverse proxy with authentication in front of it before exposing it publicly.