VS Code agents like Cline, Roo Code, and Kilo Code send large system prompts and file trees on every turn—which quickly triggers 429s on low-TPM free tiers. Pick any free provider below to inspect context limits, training privacy, and copy exact provider settings.
Choose Provider & Privacy
Model ID & Context Window
Rate Limits & Training Rule
Agent Fit (Cline / Roo / Kilo)
Trains on Free Tier
gemini-3.8-flash
Context: 1M tokens
Free tier (No Card · per-project RPM/TPM/RPD in AI Studio)
Free tier may train models (avoid for proprietary code)
Large repo context in Cline / Roo Code without hitting token truncation
No Training
openai/gpt-oss-120b
Context: 128K tokens (8K TPM free cap)
30 RPM · 1,000 RPD · 8K TPM · 200K TPD (No Card)
No training on API prompts by default
Fast Ask/Architect turns & small diffs (keep system prompts compact under 8K TPM)
⚠️ Three Gotchas When Running Cline / Roo Code on Free APIs
TPM Truncation (Groq 8K TPM): Cline's built-in tool instructions alone are several thousand tokens. Set a compact custom instruction or route large context turns to Google AI Studio (1M context).
Commercial Code Privacy: Google AI Studio Free, Mistral Experiment, and many OpenRouter :free providers may retain prompts for training. Use NVIDIA NIM, SambaNova, or paid zero-retention endpoints for client repos.
OpenRouter 50 vs 1,000 RPD: One coding task can fire 15–30 agent steps. If you use OpenRouter :free heavily, a one-time $10 balance unlock raises the daily cap from 50 RPD to 1,000 RPD.
🚀 High-TPM Fallback for Multi-File Agent Refactors
When free tiers throttle mid-diff, point Cline/Roo Code's second profile to a pay-as-you-go or self-hosted endpoint:
Novita AI Model APIs ↗: OpenAI-compatible inference across DeepSeek, Qwen, and Llama models without 8K TPM caps.