Building a Claude Code-Compatible Proxy for Databricks (with MCP, Git Tools, and Prompt Caching)

vishalveeraredd
New Contributor

‎12-03-2025 10:56 PM - edited ‎12-03-2025 11:03 PM

Many Databricks engineers have asked whether it's possible to use Claude Code CLI directly against Databricks-hosted Claude models instead of Anthropic's cloud API. This enables repo-aware AI workflows—navigation, diffs, testing, MCP tools—right inside their Databricks projects.

I recently built an open-source tool called Lynkr, which acts as a Claude Code–compatible backend that runs locally or inside a Databricks environment. The proxy forwards /v1/messages requests to Databricks Serving Endpoints, while maintaining Claude Code’s structure for chat, tools, context, and workspace actions.

GitHub Repo:
👉https://github.com/vishalveerareddy123/Lynkr

The goal is to make Databricks a first-class environment for LLM-driven development workflows—while keeping everything transparent and configurable.

What Lynkr Enables for Databricks Users

With Lynkr running locally or in a VM, you can:

✔ Use the Claude Code CLI with Databricks models

No need for Anthropic cloud access. Just point the CLI at:

 

 
 

 

export ANTHROPIC_BASE_URL=http://localhost:8080

✔ Connect to Databricks Serving Endpoints

The proxy normalizes requests into the Databricks format and returns a Claude-compatible response.

✔ Enable repo-aware intelligence

Lynkr maintains a lightweight SQLite index of your repo, including:

  • symbol search

  • cross-file references

  • framework/language detection

  • auto-generated CLAUDE.md project summary

    This feeds richer context into the model.

    ✔ Use Git + workspace tools

    The proxy implements many of the Git + tooling features you get with Claude Code:

    • status, diff, stage, commit, push

    • automated diff summaries

    • test-gating & policies

    • release-note generation

      ✔ Integrate Model Context Protocol (MCP) servers

      Lynkr automatically discovers MCP manifests (e.g., GitHub, Jira, internal tools) and exposes them as Claude Code tools.

      ✔ Use prompt caching

      You can cache repeated prompts (configurable TTL + LRU size), drastically reducing Databricks compute calls for iterative work.

      Architecture (High Level)

       

       
       

       

      Claude Code CLI
        ↓
   Lynkr Proxy
        ↓
Databricks Model Serving
        +
   Repo Indexing
        +
      MCP Tools
        +
   Git / Diff Tools

      Everything is visible and tweakable—no hidden backend logic.

      Getting Started (Databricks Setup)

      1. Install Lynkr

       

       
      npm install -g lynkr lynkr start

      2. Configure environment

      Create an .env file:

       

       

       

      MODEL_PROVIDER=databricks DATABRICKS_API_BASE=https://<your-workspace>.cloud.databricks.com DATABRICKS_API_KEY=<your-databricks-pat> WORKSPACE_ROOT=/path/to/your/repo PORT=8080 PROMPT_CACHE_ENABLED=true

      3. Point Claude Code CLI to Lynkr

       

       

       

      export ANTHROPIC_BASE_URL=http://localhost:8080/
      export ANTHROPIC_API_KEY="#dummy"

      4. Use Claude Code normally

      Commands like:

       

       

       

      claude explain file.js 
      claude diff 
      claude review
       claude apply

      will now run against Databricks models.

      Example: Rebuilding the Repo Index

      This is one tool exposed by the proxy:

       

       

      curl http://localhost:8080/v1/messages \ -H "Content-Type: application/json" \ -H "x-session-id: test" \ ...

      This refreshes CLAUDE.md, symbol search tables, and all metadata.

      Why This Matters for Databricks

      Databricks is becoming a powerful environment for:

      • LLM-assisted development

      • agent workflows

      • code automation

      • data/ETL debugging with AI

      • internal tooling built on Claude models

        Lynkr helps bridge the gap between:

        “I have a Databricks model endpoint”
        and
        “I want Claude Code-style interactions with my repo”

        without relying on a closed backend.

        Roadmap

        Upcoming features:

        • deeper LSP integration (for even smarter repo context)

        • richer diff-thread reviews

        • expanded MCP tooling

        • fine-grained Git risk scoring

        • historical test dashboards

          Links

          Dev.to 
          DeepWiki 

          Closing

          If you’re exploring AI-assisted development inside Databricks, or want to experiment with Claude tools locally, I’d love feedback. Feel free to reply here or open issues/PRs on GitHub.

           

Labels:
0 Kudos