freellmapi vs Lynkr
OpenAI-compatible proxy that stacks free tiers of 18 LLM providers (~1.7B tokens/mo) behind one /v1 — smart routing, failover, per-key quota tracking; Claude Code and Codex shims included. — versus — Self-hosted LLM gateway wrapping Claude Code, Cursor or Codex with zero code changes — strips unused tools, compresses JSON tool results ~88%, semantic-caches, tier-routes easy work to local models.
Same slot — a local proxy in front of Claude Code/Codex to cut LLM spend. lynkr's lever is compression, semantic caching and tier-routing of paid traffic; FreeLLMAPI's is routing everything to free-tier quota across providers.
| freellmapi | Lynkr | |
|---|---|---|
| Stars | 16k | 531 |
| Forks | 2.4k | 57 |
| Language | TypeScript | JavaScript |
| License | MIT | Apache-2.0 |
| Last activity | yesterday | 5 days ago |
| Topics | gateway | coding, local, gateway |
| Curated connections | 3 | 7 |
freellmapi — the curator's take
Use it for personal experimentation when you want real working inference for $0: free tiers stacked across Google/Groq/Cerebras/Mistral/etc., a Thompson-sampling router with 20-attempt failover, and Anthropic Messages + Responses API shims so Claude Code and Codex CLI run straight against the free pool. Also does embeddings, images, TTS, and an MCP server agents can query for live model availability. NOT for production or teams — single-user by design, no billing, and several provider free tiers are ToS-gray (NVIDIA's is eval-only; the repo itself says personal experimentation only). Freemium catch: new models reach free installs 30 days late unless you pay $19/yr for the live catalog.
Lynkr — the curator's take
Use it when your Claude Code/Cursor subscription limits or API bill hurt: `lynkr wrap claude` is a one-liner, and compression + routing SIMPLE-tier traffic to a free Ollama model genuinely stretches quotas. Also the escape hatch when corporate policy forces traffic through Databricks/Azure/Bedrock. NOT a model server — it only routes, so pair it with Ollama or another backend. Skip it for light usage: a proxy is one more moving part with a big config surface (tiers, budgets, cache), and semantic caching can serve stale hits on near-duplicate prompts. If you only want provider switching without the token tricks, plain LiteLLM is the boring default.