Code doesn't fix itself.Until now, it had to waitfor a human.
MACS is an open-source multi-agent system we're building to close the loop on bug fixing — read, audit, patch, verify, commit. We're in active development and looking for engineers who want to run it on real codebases.
Targeting 85%+ on SWE-bench Verified. We'll publish the numbers when we run it at full scale.
$1.52
Trillion
The estimated annual cost of poor software quality
Consortium for IT Software Quality, 2022
The bottleneck isn't finding bugs.
Static analysis tools produce reports daily. The bottleneck is fixing them — which still requires a human developer to read a report, understand the context, write a patch, test it, and commit it.
For a team of 10 engineers maintaining 500k lines of code, that report never gets fully cleared.
MACS closes that loop autonomously.
Seven stages. Any language. No config file.
Point MACS at a repo URL or drop in a zip. It handles the rest.
READ
Chunked file ingestion with dependency-aware ordering. Understands monorepos, mixed languages, and generated code.
AUDIT
Joern Code Property Graph — call/data/type-flow analysis. Finds bugs a linter cannot see: race conditions, taint paths, integer overflows, logic errors across module boundaries.
CONSENSUS
Multiple auditor agents must agree before a bug enters the fix queue. Built-in quorum prevents false positives.
FIX (Best-of-N Ensemble)
Two competing model families generate patches independently. We don't pick one model's answer. We make them compete.
ATTACK
An adversarial critic agent actively tries to break every candidate patch. Weak fixes don't survive this stage.
VERIFY (Formal Gate)
Structural diff sanity · safety pattern scan · CBMC bounded model checking · Z3 SMT constraints. Math, not vibes.
COMMIT
Verified patches are committed. Auto-PR opened. Every state transition is HMAC-SHA256 signed. Full cryptographic audit trail.
def process_data(self, items):- for item in items:- result = self.transform(item)+ for item in items or []:+ if item is not None:+ result = self.transform(item)return results
Why two models competing beats one model trying
A single LLM asked to fix a bug will fix it approximately 40–55% of the time on SWE-bench benchmarks (published literature).
MACS uses adversarial synthesis to push that ceiling higher.
We're actively measuring this at scale and will publish our numbers openly.
This transparency is a feature, not a weakness. We believe in building trust through openness.
The Rhodawk AI Product Suite
A unified autonomous DevSecOps platform. MACS fixes individual bugs with surgical precision. Rhodawk DevOps Engine is the complete intelligence layer orchestrating security research, vulnerability discovery, and autonomous remediation across entire codebases.
MACS
Multi-Agent Code Stabilizer
Autonomous bug-fixing system that reads code, audits for vulnerabilities, generates patches, and passes them through formal verification gates. Operates as the tactical execution layer.
Rhodawk DevOps Engine
Autonomous DevSecOps Control Plane v4.0
Full-stack autonomous vulnerability research and remediation system. Orchestrates reconnaissance, static/dynamic analysis, red-team attacks, and multi-model consensus for enterprise-grade security governance.
How They Work Together
RHODAWK DEVOPS ENGINE
Full intelligence layer
Routes Failed Tests & Crash Payloads To
MACS EXECUTOR
Tactical fix generation
DATA FLYWHEEL
Every fix — successful or failed — is captured as a training sample. After 50+ verified fixes accumulate, the LoRA scheduler exports proprietary fine-tuning data. Over time, the system becomes specifically optimized for your codebase's patterns and vulnerabilities.
7
Supported Languages
25+
Integrated MCP Servers
5
Proprietary Algorithms
100%
Human-Gated Submissions
Built for production
Every component designed for real-world deployment
Code Property Graph
Joern-backed causal analysis. Sees bugs that grep and lint cannot.
Formal Verification
Z3 SMT + CBMC on every patch. Not just "looks right" — mathematically gated.
Fix Memory
Committed fixes become few-shot context for future similar bugs. Improves per codebase over time.
Federated Patterns
Share abstract structural patterns between deployments. All identifiers stripped before export.
Sandboxed Execution
LLM-generated code runs in isolated Docker containers. Your filesystem is never touched by untrusted code.
Cryptographic Audit Trail
HMAC-SHA256 signed state transitions. Every decision logged and tamper-evident.
Mutation-Verified Tests
Auto-generated tests confirmed to kill real mutants, not just pass trivially.
Zero Config
Repo URL or zip file. No YAML. No annotations. No setup scripts.
MIT Licensed. Everything ships in the open.
We're not building a black box. Every agent, every pipeline stage, every verification layer is public. Fork it, audit it, deploy it yourself. If you find something wrong, open an issue.
# Clone and run MACSgit clone https://github.com/Rhodawk-AI/macs.gitcd macsdocker compose up -dShape MACS from Day One
We're inviting a small group of engineers and engineering teams to run MACS on their actual codebases during our Design Partner program. You get direct input into what we build next. We get real-world signal on what matters. It's a working relationship, not a survey.
Direct Line to Founders
You're not a ticket number.
Priority Setup
We help you deploy MACS on your stack personally.
Shape the Roadmap
Your use case influences what gets built next.
We review every application. Expect a response within 48 hours.
Built on production-grade infrastructure
No academic demos. Every component chosen for real-world deployment.
AI & Analysis
Infrastructure
These are the tools MACS is built on.
Early Supporter Program
Accelerating autonomous bug-fixing with infrastructure support
DigitalOcean
Hatch ProgramEarly supporter via Hatch startup program
Supporting the future of autonomous code stabilization and DevSecOps
Where we're going
NOW
Design Partner Program
- Real codebase runs
- Pipeline live
- Collecting signal
NEXT
SWE-bench Verified
- Full-scale evaluation
- Publish numbers openly
LATER
Managed Federation Hub
- Cross-deployment pattern sharing
FUTURE
VS Code / Zed Plugin
- IDE integration
- Inline suggestions
NOW
Design Partner Program
- Real codebase runs
- Pipeline live
- Collecting signal
NEXT
SWE-bench Verified
- Full-scale evaluation
- Publish numbers openly
LATER
Managed Federation Hub
- Cross-deployment pattern sharing
FUTURE
VS Code / Zed Plugin
- IDE integration
- Inline suggestions
We update this page as things ship. No vaporware.