How providers qualify for this coding shortlist
This comparison selects 16 of the 25 detailed providers. A provider must document an OpenAI-compatible endpoint and list at least one code-oriented model or a broadly capable Llama, Qwen, or gpt-oss family used by coding tools. That is a reproducible catalog rule, not a benchmark claim. Inclusion does not mean the model is best at code, supports every agent feature, or has enough free capacity for a large repository.
Endpoints a coding client can try
The 16 matching providers appear with their documented access type, published limits, and Base URL. The table deliberately avoids a single “best” score because agent quality depends on edit format, tool calling, context size, latency, and the codebase. Open a provider page to copy an exact model id; sending a marketing family name instead of the API id is one of the most common setup failures.
| Provider | Free access | Published limits | Card | Base URL |
|---|---|---|---|---|
| GroqCloud | Provider free tier | 30 requests per minute, 1000 requests per day | Not required | https://api.groq.com/openai/v1 |
| SambaNova Cloud | Provider free tier | 20 requests per minute, 20 requests per day | Not required | https://api.sambanova.ai/v1 |
| Cloudflare Workers AI | Provider free tier | Published in compute units | Not required | https://api.cloudflare.com/client/v4/accounts/ACCOUNT_ID/ai/v1 |
| Hugging Face Inference Providers | Provider free tier | Published as a credit balance | Not required | https://router.huggingface.co/v1 |
| SiliconFlow | Provider free tier | 1000 requests per minute | Not required | https://api.siliconflow.com/v1 |
| Fireworks AI | Provider free tier | 10 requests per minute | Not required | https://api.fireworks.ai/inference/v1 |
| Mistral La Plateforme | Provider free tier | Enforced but not published | Not required | https://api.mistral.ai/v1 |
| Alibaba Cloud Model Studio | Provider free tier | Published per model | Not required | https://dashscope-intl.aliyuncs.com/compatible-mode/v1 |
| Ollama Cloud | Provider free tier | Enforced but not published | Not required | https://ollama.com/v1 |
| Cerebras Inference | Free trial credit | 5 requests per minute | Required | https://api.cerebras.ai/v1 |
| Vercel AI Gateway | Free trial credit | Published as a credit balance | Not required | https://ai-gateway.vercel.sh/v1 |
| Together AI | Metered access | Dynamic, no fixed numbers | Not required | https://api.together.xyz/v1 |
| Nebius Token Factory | Metered access | 60 requests per minute | Not required | https://api.tokenfactory.nebius.com/v1 |
| DeepInfra | Metered access | Enforced but not published | Not required | https://api.deepinfra.com/v1/openai |
| Chutes | Metered access | Published per paid plan | Not required | https://llm.chutes.ai/v1 |
| Scaleway Generative APIs | Metered access | Enforced but not published | Required | https://api.scaleway.ai/v1 |
Match the provider to the agent loop
Autocomplete benefits from low latency and generous minute-level throughput, while repository agents need longer context and reliable tool calls. A command-line assistant may work with standard chat completions; an autonomous editor can depend on streaming, structured arguments, and repeated retries. Choose two candidates with different strengths, run the same small task, and inspect both the final patch and how many requests the loop consumed before deciding.
Free limits are amplified by agent behavior
An agent can turn one instruction into dozens of model calls as it reads files, proposes edits, runs tests, and repairs failures. A tier that feels generous in chat may hit a daily request ceiling quickly. Large context also consumes token quotas faster, and retry loops can multiply both latency and usage. Set a maximum turn count, keep commands reviewable, and use a smaller model for exploration when the client lets you split planning from final edits.
Configure without putting keys in the repository
Use the project’s client pages to generate or review the Base URL and model configuration for Codex, Cline, Continue, Cursor, or Claude Code. Keep the credential in an environment variable or the client’s own secure store. Start on a disposable branch with a narrow task, review every shell command, and require the normal test suite before accepting changes. Free access changes cost; it does not reduce the permissions an agent can exercise locally.
What was verified for this page
Provider protocol, model ids, access terms, and published limits were reviewed on 2026-07-25; no synthetic coding benchmark is implied. The methodology records how those facts qualify, while the key checker helps diagnose your own credential before a client hides the status behind a generic connection error. Recheck the provider page when a model disappears from a tool.