issue with my tokens

[smwnty]

Hello, i am a very new and inexperienced guy with this app. I am very sorry if some of the issues i have may seem too simple or stupid to answer to but please help me out. I got it running in mt terminal in about an hour maybe an hour and a half and it worked great until i wanted to actually build something. When i gave it more complicated tasks it said that the amount of tokens required is way too big and openrouter only allows me a certain amount which is below the number i need to complete the task. I only linked it to openruter and set both the amount of tokens and time to unlimited but it still did not work.In the past i have used free claude code also from github and also used openrouter and had absolutely no problems. I do not have the money to buy unlimited tokens or a big amount. Please help me out and thank you for your time!

[Erfouni]

This is almost certainly OpenRouter's 402 "insufficient credits" check, not a Vibe-Trading setting.

Before a paid request runs, OpenRouter estimates its worst-case cost: the input tokens plus the maximum number of output tokens the request allows. Vibe-Trading does not send a max_tokens value on the OpenRouter path, so OpenRouter uses its own per-request cap for that estimate. On a small or empty balance, bigger agent steps (long prompts plus tool results) no longer fit, and the request is rejected before it reaches the model. That is why simple things worked and bigger tasks fail.

Setting the key's limit to "unlimited" in the OpenRouter dashboard only removes the per-key spending cap. It doesn't add credit, so it can't fix this.

What you can do without spending money:

1. Use a free model. On https://openrouter.ai/models, filter by price = free. Free variants have ids ending in :free. Put that id in agent/.env:

   LANGCHAIN_PROVIDER=openrouter
   LANGCHAIN_MODEL_NAME=<some-model-id>:free
   

   Free models don't use credits, but they have per-minute and per-day request limits. An agent makes many calls per task, so you may hit those on long runs. Wait and retry, or split the task into smaller ones. One catch: if your OpenRouter balance is negative, even free models return 402 until it is back above zero.

2. Check what your key actually has left:

   curl https://openrouter.ai/api/v1/key -H "Authorization: Bearer $OPENROUTER_API_KEY"
   

   Look at limit_remaining and free_model_daily_requests.

3. Keep tasks smaller. Shorter instructions and fewer data-heavy steps per run mean fewer input tokens per request.

If you do add a small amount of credit later, the same error can still show up for very large single requests. The error's metadata.reason then says weight_exceeds_budget. Smaller requests or a cheaper model avoid it.

If this doesn't match what you see, please paste the exact error message (the JSON after "Error code: …"). A "maximum context length" error, for example, is a different problem with a different fix.

[smwnty]

Thank you very much for the information. What about other users? Do they use paid services for a major amount of tokens or how do they overcome this issue? Like i said in my first question i used free claude code also linked through open router and had no issues, is that because the tasks i have it were way too simple and did not require such a big amount of tokens? Are there any other ways to overcome this tokens issue like running the model locally, completely bypassing the need for tokens from openrouter and other ai services? Thank you very much for your time and have a great evening!

[Erfouni]

Glad it helped. On your questions:

Why the free Claude Code setup worked. Most likely it was using a free (:free) model, which doesn't draw on credits at all, while Vibe-Trading is set to a paid model. Compare the model ids the two use; if Vibe-Trading's doesn't end in :free, that's the difference. Task size adds to it: a finance agent step carries price tables and tool results, so each request is much bigger than a typical coding question.

What other people do. Most people running agent apps like this pay for tokens, usually a few dollars of credit, and pick a cheap model. There are a few ways around that:

• Use a subscription you already have. LANGCHAIN_PROVIDER=openai-codex signs in with a ChatGPT account (OAuth) and LANGCHAIN_PROVIDER=copilot uses GitHub Copilot, so neither is billed per token.

• Free tiers. Gemini and Groq are both supported providers and have had free, rate-limited tiers. Check their current terms. As with :free models on OpenRouter, you trade money for waiting when you hit the limits.

• Run the model locally. Yes, that works. Vibe-Trading supports Ollama, and no API key is needed:

  LANGCHAIN_PROVIDER=ollama
  LANGCHAIN_MODEL_NAME=<a model you pulled, e.g. qwen2.5:14b>
  

  Two practical warnings. First, agent tasks need a model that handles tool calls reliably over many steps, and small local models (7–8B) often don't. Something in the 14–32B range is more realistic, which means a decent GPU or a lot of RAM. Second, Ollama's default context window is small, so raise it (the OLLAMA_CONTEXT_LENGTH setting when you start ollama serve). Otherwise the long agent prompts get cut off silently. If you run Vibe-Trading in Docker, the README explains the host.docker.internal address it uses to reach Ollama on your machine.

For learning and small tasks, a :free model or a local one is fine. For heavy, multi-step analyses, a small amount of paid credit on a cheap model is usually the least frustrating option.