KodHau MCP gives your AI agent the tribal knowledge of your team: PR history, design decisions, and review comments your senior engineers never documented.
Hi Product Hunt! 👋 I'm Zhasulan, 17 y.o founder from Astana, Kazakhstan, builder of KodHau.
At 16, I led 12 developers, worked as Team Lead at a venture studio, across teams — and we used AI agents for coding, but they kept breaking our production because they had no idea what our team had already tried and rejected.
That's the knowledge problem. Tribal knowledge. Context about your codebase lives in people's heads and in discussions around the code - PRs. Not in the docs, or wikis. They get updated only when there's time.
KodHau is an MCP server that gives your AI agent access to your team's decisions, workarounds, and rejected approachesburied in years of GitHub PR history. Before your agent touches a single line of code, with KodHau it knows why the code is written that way by your engineers.
The proof: I used KodHau to fix an 8-month-old bug in Microsoft's .NET runtime, their flagship repo. Someone else tried fixing that issue already — 200 lines of code, wrong approach, abandoned. Our fix was 7 lines. KodHau found decisions Microsoft engineers made 4 years ago for this fix. And the same applies for ANY repo.
2-minute setup. Works with Cursor, Claude Code, any MCP client.
Happy to answer questions about MCP, GitHub API, or how tribal knowledge injection works 🚀
@zhas_srk under the hood, how are you handling token limits for massive repos? is there a local vector db filtering the noise before the mcp sends it to cursor?
@mohsinproduct KodHau pulls PR history directly from GitHub. Before your Cursor or Claude Code touches a line of code, the MCP finds PRs by exact file path and function name - so you get the PRs where your engineers actually wrote about that specific code, not semantically similar ones. No vector DB, no RAG noise, no data stored on our side. GitHub -> straight to your agent's context window.
Report
Congrats on the launch bro
PR history as memory/context for AI agents is actually such a smart idea. Most AI coding tools understand the code, but not the reasoning behind the code or why certain approaches were rejected
The knowledge problem is the biggest bottleneck for AI right now. I’m tired of Claude Code suggesting a refactor that we already tried and rejected three months ago in a PR. 😅 Having an MCP server that actually looks at why decisions were made is a massive unlock for team productivity. Support on the launch, @zhas_srk
@vikramp7470 thank you Vikram! That's the pain point I faced myself. So before KodHau touches any code, it pulls the PR where your team debated and rejected that refactor - the reasoning your senior engineer never wrote documented.
A lot of teams worry about prompt-injection and privilege leakage once agents start reading Issues/PR text—what design choices in KodHau reduce that risk, and what are the recommended permission/scoping defaults for GitHub tokens?
@curiouskitty Great question. KodHau uses a GitHub App rather than personal tokens - scopes are fixed at the app level to Issues (read-only) and Metadata (read-only), so there's no way to accidentally over-permission. The scopes are visible for users. During installation users choose exactly which repositories to grant access to so you can scope it to just the repos your agent works on, and change it any time.
On prompt injection - the MCP surfaces all extracted context to the user before injection, so you see exactly what goes in (decisions, constraints rejected approaches), you're reading your own repo's history, not external input. Solution is - a sanitization layer before injection.
Report
Something we started doing a couple months back is we added a line in our Agents.md, that instructs our agents to add a changelog line for every change that it makes, every commit, every PR, including info like reason, changes and relevant context.
Over time we’ve built a knowledge base that now allows our agents to instantly get info about similar prior issues.
Something like that would be a great addition.
@arag_agrawal Smart approach for smaller repos. Two scaling problems though: on large enterprise codebases (think 5+ years, 15k+ PRs) that AGENTS.md becomes enormous fast - and every agent session loads it entirely, burning context tokens on history that's irrelevant to the current task. On top of the context window problem - .md files only work if someone remembers to update it and if the agent is explicitly told to read it.
KodHau flips this: instead of loading everything upfront, it pulls only the PRs relevant to the specific files your agent is touching right now. On-demand context vs. always-on context.
Report
@zhas_srk makes sense! Congrats on the launch btw, cool stuff
The .NET runtime example is convincing — 7 lines vs 200 because the agent knew what was already tried is exactly the kind of concrete proof that cuts through AI hype. Building multi-agent systems myself, the biggest silent failure mode is agents confidently repeating approaches that were already rejected. Curious how KodHau handles repos where PR descriptions are low quality or just 'fix bug' — does the tribal knowledge layer degrade gracefully or does it need a minimum bar of PR hygiene to be useful?
Replies
KodHau: Tribal Knowledge for AI Agents
Hi Product Hunt! 👋 I'm Zhasulan, 17 y.o founder from Astana, Kazakhstan, builder of KodHau.
At 16, I led 12 developers, worked as Team Lead at a venture studio, across teams — and we used AI agents for coding, but they kept breaking our production because they had no idea what our team had already tried and rejected.
That's the knowledge problem. Tribal knowledge. Context about your codebase lives in people's heads and in discussions around the code - PRs. Not in the docs, or wikis. They get updated only when there's time.
KodHau is an MCP server that gives your AI agent access to your team's decisions, workarounds, and rejected approaches buried in years of GitHub PR history. Before your agent touches a single line of code, with KodHau it knows why the code is written that way by your engineers.
The proof: I used KodHau to fix an 8-month-old bug in Microsoft's .NET runtime, their flagship repo. Someone else tried fixing that issue already — 200 lines of code, wrong approach, abandoned. Our fix was 7 lines. KodHau found decisions Microsoft engineers made 4 years ago for this fix. And the same applies for ANY repo.
2-minute setup. Works with Cursor, Claude Code, any MCP client.
Happy to answer questions about MCP, GitHub API, or how tribal knowledge injection works 🚀
PicWish
@zhas_srk under the hood, how are you handling token limits for massive repos? is there a local vector db filtering the noise before the mcp sends it to cursor?
KodHau: Tribal Knowledge for AI Agents
@mohsinproduct KodHau pulls PR history directly from GitHub. Before your Cursor or Claude Code touches a line of code, the MCP finds PRs by exact file path and function name - so you get the PRs where your engineers actually wrote about that specific code, not semantically similar ones. No vector DB, no RAG noise, no data stored on our side. GitHub -> straight to your agent's context window.
Congrats on the launch bro
PR history as memory/context for AI agents is actually such a smart idea. Most AI coding tools understand the code, but not the reasoning behind the code or why certain approaches were rejected
Rooting for this one fr
KodHau: Tribal Knowledge for AI Agents
@codewithriza Thank you so much!🙏 That's exactly it — the current code is the what, while discussions of engineers around the code is the why
nFactorial AI
Congrats on the Launch, Zhasulan!
KodHau: Tribal Knowledge for AI Agents
@yuriy_kimm thank you, Yuriy!
The knowledge problem is the biggest bottleneck for AI right now. I’m tired of Claude Code suggesting a refactor that we already tried and rejected three months ago in a PR. 😅 Having an MCP server that actually looks at why decisions were made is a massive unlock for team productivity. Support on the launch, @zhas_srk
KodHau: Tribal Knowledge for AI Agents
@vikramp7470 thank you Vikram! That's the pain point I faced myself. So before KodHau touches any code, it pulls the PR where your team debated and rejected that refactor - the reasoning your senior engineer never wrote documented.
Product Hunt
KodHau: Tribal Knowledge for AI Agents
@curiouskitty Great question. KodHau uses a GitHub App rather than personal tokens - scopes are fixed at the app level to Issues (read-only) and Metadata (read-only), so there's no way to accidentally over-permission. The scopes are visible for users. During installation users choose exactly which repositories to grant access to so you can scope it to just the repos your agent works on, and change it any time.
On prompt injection - the MCP surfaces all extracted context to the user before injection, so you see exactly what goes in (decisions, constraints rejected approaches), you're reading your own repo's history, not external input. Solution is - a sanitization layer before injection.
KodHau: Tribal Knowledge for AI Agents
@arag_agrawal Smart approach for smaller repos. Two scaling problems though: on large enterprise codebases (think 5+ years, 15k+ PRs) that AGENTS.md becomes enormous fast - and every agent session loads it entirely, burning context tokens on history that's irrelevant to the current task. On top of the context window problem - .md files only work if someone remembers to update it and if the agent is explicitly told to read it.
KodHau flips this: instead of loading everything upfront, it pulls only the PRs relevant to the specific files your agent is touching right now. On-demand context vs. always-on context.
@zhas_srk makes sense! Congrats on the launch btw, cool stuff
KodHau: Tribal Knowledge for AI Agents
@arag_agrawal thanks Arag! try it out on kodhau.com :D
nFactorial AI
Congrats on the launch, Zhasulan!
Very impressive!
KodHau: Tribal Knowledge for AI Agents
@suleimenov thank you Arman! 🙏
The .NET runtime example is convincing — 7 lines vs 200 because the agent knew what was already tried is exactly the kind of concrete proof that cuts through AI hype. Building multi-agent systems myself, the biggest silent failure mode is agents confidently repeating approaches that were already rejected. Curious how KodHau handles repos where PR descriptions are low quality or just 'fix bug' — does the tribal knowledge layer degrade gracefully or does it need a minimum bar of PR hygiene to be useful?