BugForge AI

Inspiration

Bug reports often start with something simple: a crash, a failing test, or an unexpected reproduction. The difficult part is turning that raw evidence into a useful investigation.

I have been building BugForge around real crash and reproduction workflows, and I wanted to make those capabilities accessible through a natural-language agent instead of requiring developers to manually inspect logs, stack traces, fingerprints, and reports.

That inspired BugForge AI: an agentic bug-investigation platform that connects my crash-analysis engine to a self-hosted MCP server for the Alexa+ track.

What it does

BugForge AI lets an agent investigate real crash and reproduction data through MCP tools.

A user can ask:

«“Investigate my latest reproduction and tell me what happened.”»

The agent can use BugForge to:

Find available reproductions
Read crash artifacts
Parse stack traces
Analyze crash indicators
Generate deterministic fingerprints
Compare investigation evidence
Inspect bounded source-code ranges
Generate structured reports
Check project and reproduction status

The Alexa+ experience turns these capabilities into a conversational workflow.

For example:

User ↓ “Investigate my latest crash” ↓ Alexa+ ↓ BugForge MCP ↓ list_reproductions ↓ get_reproduction ↓ analyze_crash ↓ get_fingerprint ↓ generate_report ↓ Evidence-based response

The agent operates on real BugForge analysis and reproduction data, rather than returning hard-coded demo responses.

BugForge AI also includes an optional Amazon Bedrock + Strands Agents integration for the AWS Builder challenge.

How we built it

I built BugForge AI primarily in Python and extended my existing BugForge crash-analysis workflow with an agentic layer.

The architecture is:

Alexa+ / MCP Client │ Streamable HTTP │ ▼ BugForge MCP Server │ ┌────┼────┐ ▼ ▼ ▼ Reproduction Crash Source Tools Tools Tools │ ▼ BugForge Engine │ ┌────┼────┐ ▼ ▼ ▼ Parser Fingerprint Reporting │ ▼ Evidence-based result

I exposed real BugForge functionality through MCP tools including:

list_reproductions get_reproduction analyze_crash parse_stacktrace inspect_source get_fingerprint generate_report get_project_status

For the AWS Builder challenge, I added a Strands Agents integration that can use Amazon Bedrock while keeping my deterministic BugForge analysis engine as the source of technical evidence.

The AI layer orchestrates tools and explains evidence rather than pretending that the model itself discovered or reproduced a crash.

Challenges we ran into

The biggest challenge was building the project completely solo while extending an existing crash-analysis system and implementing the Alexa+ integration at the same time.

I wanted the Alexa+ component to be a real integration rather than a chatbot mockup.

Another challenge was deciding how much information each MCP tool should expose. An agent needs enough information to investigate a reproduction, but unrestricted filesystem access would be unsafe for a network-accessible service. I therefore designed explicit investigation operations and bounded source inspection.

I also had to preserve the reliability of my existing BugForge pipeline while adding an AI layer on top of it.

Crash data can vary significantly between runtimes, sanitizers, stack formats, and failure types, so the system also needs to work with imperfect real-world evidence.

Accomplishments that we're proud of

I'm proud that I turned my existing BugForge crash-analysis project into an agent-ready investigation platform instead of building a disconnected hackathon prototype.

I'm especially proud of building a real self-hosted MCP server using Streamable HTTP and connecting it to actual BugForge investigation capabilities.

I built real MCP tools for crash analysis, reproduction discovery, stack-trace parsing, deterministic fingerprinting, source inspection, and report generation.

I'm also proud of connecting the project to the Alexa+ agentic workflow while keeping the underlying technical evidence grounded in BugForge's real analysis engine.

I added an optional Amazon Bedrock + Strands Agents integration, allowing the project to participate in the AWS Builder challenge.

Most importantly, I built the entire project as a solo developer, bringing together the crash-analysis engine, MCP, Alexa+, and AWS agent capabilities into one system.

What we learned

I learned that an AI agent becomes much more useful when it has access to specialized tools that produce deterministic evidence.

Instead of asking an AI model to guess the cause of a crash, I give the agent real operations for collecting and analyzing evidence.

I learned how to: