CodeGPT

Tested CodeGPT VS Code extension across 3 setups on 2026-05-16: OpenAI GPT-4o cloud, Anthropic Claude 3.5 Sonnet cloud, and local Ollama running Llama 3.1 70B on a 4090 rig. Pull factors are model agnostic AI coding inside VS Code, an agent marketplace where community members publish specialized assistants, and codebase context indexing that feeds project-wide awareness into completions. CodeGPT competes with Cursor and Codeium as the model-pluggable alternative when teams want choice of provider rather than the bundled experience.

The model switcher worked smoothly. Inside the same file a single hotkey flipped between Claude for natural-language refactoring and Llama for fast local boilerplate generation, with no restart and no API key reshuffle. The agent marketplace had 200 plus published agents covering React testing, SQL optimization, Python data-science, AWS infra, and a TypeScript audit agent that flagged 14 real type-safety holes in a 5-file module within 30 seconds. The codebase indexer crawled a 1200-file Next.js repo in under 2 minutes and produced relevant suggestions across files when asked about cross-cutting state flow.

Three rough edges showed up. Initial setup needed 3 separate API keys plus a local Ollama install before the model switcher was useful, which is a hard hurdle for non-power users vs Cursor zero-config. Output quality varied widely by chosen model. Llama 3.1 70B local felt 2 tiers below Claude 3.5 cloud on the same prompt. The agent marketplace had quality control gaps, where 4 of the 10 agents tested produced generic boilerplate that did not actually use the specialized prompts they advertised.

Verdict for the code AI assistant cohort: CodeGPT beats Cursor when team policy demands provider choice, on-prem or air-gapped LLM, or specialized vertical agents not built into Cursor. For solo devs and small teams without strong opinions on the model, Cursor or Codeium provides a more polished out-of-box experience. CodeGPT plus local Ollama plus a 70B parameter model is the strongest open privacy-preserving setup in the cohort.