Codivity

Github powered intelligence layer for coding agents and humans

Comment

Codivity

GitHub-powered intelligence layer for coding agents and humans

Inspiration

Modern repositories — especially fast-moving ones — accumulate enormous institutional knowledge inside issues and pull requests. Architectural debates, rejected approaches, edge cases, migration strategies — they’re all there, but scattered.

At the same time:

  • Engineering teams are seeing higher PR volumes, partly due to AI-assisted coding
  • New contributors (human or AI) lack historical context
  • Coding agents can generate code, but not repository memory

Whether in open source or private organizations, teams struggle with the same problem:

How do we turn GitHub history into structured intelligence that both humans and coding agents can rely on?

Codivity was built to become that shared intelligence layer.

What it does

Codivity indexes GitHub issues and pull requests from linked repositories and converts them into a searchable, reasoning-ready knowledge base.

It enables:

For Humans

  • Ask natural language questions about past discussions and decisions
  • Understand why certain implementations were accepted or rejected
  • Detect duplicate or similar PRs
  • Onboard faster by querying repository history directly

For Coding Agents

  • Retrieve grounded repository context before implementing new features
  • Reference past implementation patterns and architectural decisions
  • Avoid repeating rejected strategies
  • Generate changes aligned with project conventions and prior discussions

Codivity acts as a persistent memory layer — powering both maintainers and autonomous coding agents with contextual awareness.

How we built it

We built Codivity using the Elastic platform as the backbone of indexing, retrieval, and agent exposure.

1. GitHub Ingestion & Indexing

  • Ingest issues, PR descriptions, discussions, and metadata from linked repositories
  • Structure and index this data in Elasticsearch
  • Design the indexing pipeline to support both private and public repositories

Elastic’s indexing capabilities allowed us to quickly build a scalable and flexible knowledge store.

2. Retrieval + Multi-LLM Reasoning

  • Use Elasticsearch for structured and semantic search
  • Retrieve relevant threads before invoking reasoning
  • Leverage Elastic’s support for multiple LLM providers
  • Ensure responses are grounded strictly in retrieved repository context

This keeps the system retrieval-first rather than model-first.

3. Agent Integration via MCP

  • Expose Codivity as an MCP-compatible agent
  • Allow external coding agents to query repository intelligence programmatically
  • Import tools via MCP for extended workflows
  • Use Elastic’s REST APIs for automation and integration into development pipelines

This makes Codivity usable both as a user-facing interface and as backend infrastructure for other agents.

Challenges we ran into

  • Long PR threads exceeding LLM context limits
  • Noisy discussions with mixed signal quality
  • Designing chunking and ranking strategies that preserve decision integrity
  • Preventing hallucinations through strict retrieval grounding
  • Building abstractions that support multiple LLMs consistently

Balancing flexibility for private repositories while maintaining strong retrieval precision was a key engineering challenge.

Accomplishments that we're proud of

  • Building an end-to-end GitHub → Elasticsearch → Agent pipeline
  • Creating a shared intelligence layer usable by both humans and coding agents
  • Successfully exposing the system via MCP for interoperability
  • Designing the system to remain LLM-provider agnostic
  • Demonstrating Elastic’s strengths in indexing, retrieval, and agent infrastructure

We’re particularly proud that Codivity is not just a Q&A tool — it is infrastructure for repository-aware automation.

What we learned

  • Repository history is a high-value, underutilized dataset
  • Retrieval quality is foundational to trustworthy agent behavior
  • Humans and coding agents benefit from the same structured memory layer
  • Elastic’s ecosystem significantly reduces friction in building modular AI systems
  • Interoperability (via MCP and APIs) is critical for agent-native tooling

We learned that intelligence in software development isn’t just about generating code — it’s about understanding context.

What's next for Codivity

  • Advanced PR similarity detection for automated triage
  • Repository-level decision summarization dashboards
  • CI/CD workflow integrations
  • Support for additional sources like commit history and documentation
  • Deeper coding-agent integrations for feature implementation workflows

Our long-term vision is to make every repository — private or public — agent-ready and context-aware by transforming GitHub activity into a durable intelligence layer.

Built With

Share this project:

Updates