Video summary

I Ran Claude Code for FREE… Here's How

Main summary

Key takeaways

Technology

Tech summary (Claude Code + local models via LM Studio)

  • Claude Code now can use local models: Instead of always relying on Anthropic’s cloud, Claude Code can talk to an LLM running on the user’s machine (“run an LLM right on my laptop”).
  • LM Studio adds support for Claude Code: LM Studio is highlighted as a graphical local model host that can serve models to Claude Code.
  • Question being tested: “Is it worth it?” The video compares local-model performance against a “gold standard” Anthropic Claude Opus 4.5 (the cloud model), using a real coding task.

Product/tools mentioned & key features

LM Studio

  • Cross-platform: macOS, Windows, Linux.
  • Model download/search UI.
  • Supports different model formats depending on hardware:
    • MLX: Apple-silicon only (better performance on macOS).
    • GGUF: works across platforms (more portable, useful for moving to other machines).
  • Models are quantized (examples shown: 4-bit versions) to make them feasible to run locally.

Merlin AI (discount + workflow features)

  • “All-in-one” assistant that consolidates ChatGPT, Claude, Gemini, etc.
  • Features shown:
    • Summarize what you’re reading directly from the webpage.
    • Deep research mode generating structured reports from multiple sources.
    • Quick modes like web, academic, Reddit search.
  • Review/analysis angle: argues it’s cheaper than paying for each provider separately because it buys API access in bulk.
  • Includes a promo code reducing the annual total to $60 (~$5/month) (deal time-limited).

Guide/tut workflow: running Claude Code with local hosting

  1. Install LM Studio and confirm the update that adds Claude Code support.
  2. Choose/download a suitable model
    • Start small first to validate behavior.
    • Model quality isn’t guaranteed: some models worked poorly with Claude Code (the creator says a 1.2B model “didn’t really work so well”).
  3. Configure context + resources
    • The creator increases context length (mentions “maximum” / huge context values).
    • Hardware considerations:
      • Example: 128 GB unified memory allows bigger models.
      • On smaller systems (e.g., MacBook Air with limited unified memory), use smaller models.
  4. Install/configure Claude Code
    • Uses Claude Code docs “quick start guide” with a curl command (PowerShell alternative on Windows).
    • Claude Code can be launched from within a project directory.
  5. Host the local model in LM Studio
    • Use LM Studio’s Developer tab to load/serve the model.
    • Note the server address (IP) and port (commonly port 1234 is mentioned).
  6. Connect Claude Code to the local server
    • Create/edit Claude Code settings in ~/.Claude/ (VS Code editing shown).
    • Create a custom JSON settings file like LM Studio.settings.json containing:
      • base URL pointing to LM Studio’s IP + port
      • Anthropic auth token not required (unless you enable it)
      • model can be left as default
    • Launch Claude Code with:
      • --settings pointing to that JSON file.
  7. Observe logs and prompt processing
    • Developer logs show Claude Code sending large context (prompt processing is emphasized).

Main analysis/results from the model comparisons

Cloud baseline (Anthropic Opus 4.5)

  • Treated as a “golden standard.”
  • Claude Code successfully updated dependencies (e.g., updating package.json and package-lock.json), then things worked.

Local model test #1: “GPT-OSS 20B”

  • Result: “kind of a failure”
  • It updated only the lockfile (not package.json) and didn’t fully achieve the task as expected.
  • Conclusion: not all models behave equally for the same Claude Code task.

Local model test #2: “GPT-OSS 120B”

  • Result: improved success
  • It updated dependencies (including package.json), and the updated app ran.
  • Notes on performance:
    • Prompt processing takes a while and memory usage spikes due to huge context.
    • Example insight: memory pressure rises significantly when context is included, even if the model is “only” a certain disk size.

Key technical takeaway

  • The video repeatedly emphasizes that prompt processing time and resource usage matter a lot, and vary by hardware.
  • It also cites general benchmark behavior: some systems are fast at prompt processing while others excel at token generation (DGX Spark vs Apple M4 Max/M3 Ultra is referenced).

Main speakers/sources

  • Speaker: The video creator (no name provided in the subtitles; they narrate the setup, testing, and conclusions).
  • Sources/tools:
    • Anthropic Claude Code / Opus 4.5
    • LM Studio
    • Merlin AI
    • Claude Code documentation / quick start guide

Original video