This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
What I Built
An AI coding agent can finish a task in minutes. Reviewing what it wrote can take an hour, and you still can't be sure you caught everything.
That's the new bottleneck. Agents like Claude Code, Cursor, and Copilot write thousands of lines at a time, but a human still has to read them. So we skim, or we click "accept" and hope. Then the agent turns out to have over-engineered a simple feature, stubbed out the logic that mattered, or quietly rewritten the part of the project that was working.
CodeStream is the reviewer that reads everything first, then tells you what to worry about.
Point it at a GitHub repo, branch, or commit range, and it produces:
- 🔍 A risk-ranked review: high, medium, and low findings, each tied to specific files
- ✅ A "Before You Trust It" checklist: the exact files and behaviours to verify by hand
- 🎙️ A narrated audio briefing: a short podcast of what changed and why it matters, for your commute or a walk
- 💬 Cited Q&A: ask "where does session state live?" and get an answer that points at real files
Who would pay: developers and small teams who ship code written by AI agents, and freelancers who have to stand behind agent-written code they hand to clients. For them, one missed regression costs far more than a review tool.
Where it is today: CodeStream is a proof-of-concept MVP, free to try with no credit card. I'm keeping it free while I learn what people actually use. The plan is to charge for what costs me money to run: private repositories, larger codebases, and unlimited audio briefings. Those are where the model tokens and ElevenLabs characters go.
Demo
🔗 Live app: https://codestream-web.onrender.com
Code
Tech-Vexy
/
CodeStream
Helps you review changes made by an Agent to your codebase
CodeStream
Know what your coding agent actually built.
Codebase tours, commit-change reviews, narrated audio briefings, and interactive Q&A.
Overview
AI coding agents (such as Claude Code, Cursor, GitHub Copilot, Codex, and Devin) can produce hundreds or thousands of lines of code in minutes. Reviewing these changes line-by-line is exhausting, while merging them blindly invites architectural drift, security vulnerabilities, stubbed logic, and regressions.
CodeStream solves this by acting as your automated tech lead and reviewer:
- Digests repositories and commit diffs using intelligent file filtering, budget packing, and sensitive-file isolation.
- Analyzes code structure and risks with specialized multi-agent AI pipelines powered by Google Gemma 4 on Mastra.
- Generates executive reports detailing architecture, component maps, data flows, reading orders, and severity-ranked risk audits.
- Produces podcast-style audio briefings ("casts") voiced by ElevenLabs, streamable with scrubbing support directly in your browser.
- Enables interactive grounded Q&A backed by Voyage AI vector embeddings, letting you…
How I Built It
1. Ingest, without sending secrets anywhere. CodeStream pulls the repo or diff from GitHub and filters out noise: lockfiles, binaries, node_modules, build output. Credential files (.env, private keys, .pem, service-account JSON) are detected by name and never sent to the model. They appear in the report as findings instead. Source is capped at 600k characters per run and diffs at 400k, so cost stays predictable.
2. Durable workflows with Temporal. An analysis can run for minutes and call several external APIs. Each stage (ingest, chunk, analyse, narrate, embed) is a Temporal activity with exponential backoff, so a rate limit pauses the job instead of killing it.
3. Three agents on Gemma, orchestrated by Mastra. A Codebase Analyst maps architecture and components, a Change Reviewer reads diffs for risk, and a Narrator rewrites the findings as a script meant to be heard, with no raw code or long file paths. All three run on Gemma 4 31B through the Google AI API, with outputs validated by Zod schemas. Large inputs are split into chunks, summarised in parallel at limited concurrency, then synthesised.
4. Voice with ElevenLabs. The script is synthesised in segments, with context carried across boundaries so the voice doesn't reset mid-thought. The MP3 is stored in MongoDB GridFS and streamed with HTTP 206 range requests, so scrubbing is instant.
5. Grounded Q&A. Code chunks are embedded with Voyage AI (voyage-code-3) and stored in MongoDB. Questions retrieve the most relevant passages, and the model is constrained to cite real files.
6. Render. One render.yaml blueprint deploys the web service and the worker.
Building it with agents. I treated the build as a test of the problem CodeStream solves. I started with Backboard's open-source terminal agent, R-CLI, to scaffold the project and write the initial plan. When my R-CLI credits ran out, I handed the remaining work to GitHub Copilot. Two different agents wrote most of this code, which is exactly why I needed something to review it. (The app was called CodeCast in the first session, before I renamed it.)
Next: self-hosted Gemma for teams that can't send code out, and a GitHub App that posts the review checklist directly on PRs.
Why Does Open Innovation Matter?
A code reviewer has a hard trust problem: it reads your whole codebase. And the cost model is unusual, because one review can mean hundreds of thousands of characters of input.
Building on open-weight Gemma changes both:
- Cost I can plan around. The same open weights are available from several providers, so I'm not tied to one vendor's pricing as repos get bigger.
- A privacy path for sensitive code. Teams that can't send source to a third party can run the same model on their own infrastructure. Today CodeStream calls Gemma through Google AI. Self-hosting is the roadmap for those customers.
- A stack you can inspect. Gemma, Mastra, and Temporal are open, so a customer can audit how their code is analysed. For a tool whose pitch is "don't trust what you haven't verified," that matters.
My Agent Session
I started the build on Oct 16 in Backboard's R-CLI. In about three hours it researched the current Mastra, Temporal, ElevenLabs, Gemma, and Sentry docs, planned the work as a ten-step checklist, and wrote the first modules: GitHub ingestion, file filtering, chunking, the MongoDB/GridFS store, the Temporal workflow and worker, and the API. It was about to run its first typecheck when my credits ran out, so I carried the repo to GitHub Copilot.
My First User
Honestly, that's me. CodeStream is a proof of concept and I'm its only user so far, so treat this as dogfooding, not traction. Nobody has paid yet. If you try it on your own agent-written PRs, I'd like to hear what it misses.
Prize Categories
- Best Use of Gemma: Gemma 4 31B powers all three agents.
- Best Use of Render: the web service and worker deploy from a single blueprint.
- Best Use of Mastra: Mastra orchestrates the multi-agent pipeline with structured outputs.
- Best Use of ElevenLabs: every review is narrated and streamed with scrubbing.
- Best Use of MongoDB Atlas: casts, chunks, embeddings, and audio all live in MongoDB.
- Best Use of Temporal: durable workflows ride out rate limits and provider failures.
- Best Use of GitHub Copilot: Copilot completed the build after the initial scaffold.
- Best Use of Backboard: R-CLI scaffolded the project and produced the initial plan.
Top comments (0)