ИИ Агенты Работают Бесплатно? (Без RAM и GPU)

Core idea

You can run AI agents without paying per-input fees by routing requests to free hosted models (via OpenRouter) instead of paid commercial models or self-hosting large models.

• Trade-offs: free models reduce direct cost but are lower-quality, rate-limited, and less reliable under sustained load.
• Self-hosting small models still requires substantial CPU/GPU and RAM, which carries its own cost and complexity.

Key takeaway: free hosted models are useful for experimentation and light agent tasks, not for production-grade, high-throughput, or high-quality use cases.

Key platforms and technologies

• OpenRouter
  • A hub that hosts and benchmarks many free models and provides an API.
  • Offers a “Free Models Router” that rotates across available free models to improve availability.
• Agent UI / agent repository
  • Example agent project (cloneable from GitHub) that runs locally (Docker Compose) and can use OpenRouter as the model backend.
• CLD-code (coding/agent framework)
  • Can be configured to route model calls through OpenRouter (used for coding agents).
• Development tools
  • VS Code (or other editors), Docker / Docker Compose for local deployment.
• Mentioned model examples and providers
  • Free model examples surfaced in OpenRouter’s benchmarks (e.g., “stepfun”-like names).
  • References to Anthropic and OpenAI/open-source GPT models (access may be restricted).

Practical walkthrough (high-level steps)

1. Clone the agent repo from GitHub into a local folder and open it in VS Code (or any IDE).
2. Start the agent with Docker Compose:
   • Ensure any example env file is renamed (remove the “example” suffix if present).
   • Run: docker compose up (initial image/container creation may take a few minutes).
3. Open the local web UI, create an account/project, and connect a model:
   • Create an OpenRouter API key: OpenRouter → Settings → API keys → create.
   • Paste the API key into the agent UI and select a model (for example, one of the free models listed on OpenRouter).
4. Use the Free Models Router option to let OpenRouter rotate among free models automatically (improves availability but obscures which specific model responds).
5. Connect CLD-code to OpenRouter:
   • Follow OpenRouter docs → Guides → “Coding agents” (CLD-code page).
   • Create the required local settings file (e.g., settings/local.json) and paste the configuration object from the OpenRouter docs.
   • Replace provider-specific fields as instructed: insert your OpenRouter API key and set the model identifier.
   • Restart CLD-code so it picks up the new API/model configuration.
6. Test by sending prompts through the agent/CLD-code; responses routed via OpenRouter’s free models should not incur per-input charges.

Benchmarks and behavior

• OpenRouter exposes benchmarks and ranks models for “agent tasks” (tool usage, calling tools, processing results). Some free models score well on these tasks.
• The Free Models Router rotates models if a listed model is unavailable; this provides basic functionality but can lead to inconsistency in responses.
• Attempting to use large open-source models via OpenRouter can fail if access is restricted; you may see “access limited or impossible” errors.

Limitations, pitfalls, and troubleshooting

• Hardware: self-hosting even “small” performant models requires substantial CPU/GPU and RAM.
• Performance & reliability: hosted free models are suitable for experimentation and light agent tasks, but expect:
  • Lower response quality
  • Rate limits
  • Intermittent availability
• Common setup errors:
  • Wrong filenames (e.g., not creating settings/local.json or leaving .env/example unchanged)
  • Typos in API keys or model IDs
  • Forgetting to restart services after configuration changes
• If a chosen model is unavailable or requires special access, you’ll encounter errors; rotating to other free models may help but will reduce consistency.

Guides, tutorials, and resources

• OpenRouter docs → Guides → “Coding agents” (CLD-code integration guide)
• Repository walkthrough for the example agent project (clone + Docker Compose + local UI)
• Steps to create an OpenRouter API key and select a model from OpenRouter’s model list
• CLD-code configuration instructions (create settings/local.json and insert OpenRouter config)
• The original video also referenced a paid masterclass for deeper CLD-code learning (promotional)

Main speakers and sources referenced

• OpenRouter — platform, docs, Free Models Router, model benchmarks
• CLD-code — coding agents integration
• Local dev tools demonstrated: Docker / Docker Compose, VS Code
• Mentions/comparisons: Anthropic, OpenAI / open-source GPT models
• Video narrator / content creator — presenter who performed the demo and walkthrough (unnamed)