djlougen/busybee-cpu

★ 21PythonAudience · developerComplexity · 3/5LicenseSetup · moderate

Mindmap

mindmap
  root((busybee-cpu))
    What it does
      Cuts LLM API calls
      Local decision routing
    Classifier
      Ensemble of 3 models
      SGD Naive Bayes Logistic
      30ms response time
    Actions classified
      Read file
      Run tests
      Apply patch
      Escalate to LLM
    Integration
      HermesAgent-20
      Acts as model endpoint
    Performance
      96.4% accuracy
      20 of 20 real scenarios

mindmap root((busybee-cpu)) What it does Cuts LLM API calls Local decision routing Classifier Ensemble of 3 models SGD Naive Bayes Logistic 30ms response time Actions classified Read file Run tests Apply patch Escalate to LLM Integration HermesAgent-20 Acts as model endpoint Performance 96.4% accuracy 20 of 20 real scenarios

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Reduce AI API costs and latency in an agentic coding pipeline by routing repetitive file-read and test-run decisions to a local CPU classifier.

USE CASE 2

Integrate a drop-in local model endpoint with HermesAgent-20 that handles routine steps without any network calls.

USE CASE 3

Train and deploy a 30ms routing classifier on your own hardware to make an AI agent faster and cheaper at 96.4% accuracy.

USE CASE 4

Add a safety escalation valve so only genuinely complex decisions reach the cloud language model, cutting unnecessary API spend.

Tech stack

Pythonscikit-learn

Getting it running

Difficulty · moderate Time to first run · 30min

Requires cloning the repo, installing Python dependencies, and running a training command before the local server endpoint can start.

MIT license, use freely for any purpose, including commercially, as long as you keep the copyright notice.

In plain English

busyBee-cpu is a small Python tool that reduces how often an AI language model needs to be called while an AI agent is running tasks. The core insight is that many steps in an agent's work are obvious and mechanical: before editing a file, you need to read it first, after making a change, you should run the tests. These decisions do not require any real thinking, yet a typical agent setup calls the language model to make them anyway, which wastes time and money. The tool trains a lightweight classifier that runs on ordinary CPU hardware and handles these repetitive routing decisions on its own. When the agent reaches a decision point, the classifier picks from four possible next actions: read a file, run tests, apply a patch, or escalate to the language model for actual reasoning. The escalate option is the safety valve, if the classifier is not confident or the situation is genuinely complex, it hands off to the language model. The classifier itself takes about 30 milliseconds to respond and never calls any external API. The classifier is built from an ensemble of three simple statistical models (SGD, Naive Bayes, and Logistic Regression) voting together. It was trained on 819 examples and evaluated on about 12,000 real-world software engineering problems it had never seen, achieving 96.4% accuracy in choosing the correct next action. The project reports a score of 20 out of 20 on a separate set of real-world agent scenarios. The primary integration target described in the README is HermesAgent-20, an AI agent framework. The adapter makes the classifier look like a standard language model endpoint so the agent system treats it as just another model provider, but internally it runs local Python classifiers instead of calling a cloud API. Setup involves cloning the repository, installing Python dependencies, training a model with a provided command, and starting a local server. The project is released under an MIT license.

Copy-paste prompts

Prompt 1

I am using HermesAgent-20 and want to reduce API calls. Walk me through setting up busybee-cpu as a local model endpoint so it handles routine decisions like reading files and running tests.

Prompt 2

Help me train the busybee-cpu classifier on my own dataset of agent decision steps and evaluate its accuracy before deploying it as an endpoint.

Prompt 3

How does busybee-cpu decide when to escalate to the language model instead of routing locally? Explain the confidence threshold and the escalate action logic.

Prompt 4

I want to extend busybee-cpu to support a fifth action type beyond the current four. Show me how to add a new class to the classifier and retrain it.

Open on GitHub → Explain another repo

← djlougen on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.