Skip to content

Shared tools package

The repository root installs the agentic-tools Python package from tools/. It contains model clients, discovery utilities, local-model helpers, and benchmark code shared by the other packages.

agentic-tools and the agentic_v2 runtime have separate provider-routing implementations. A model setting that affects one does not automatically affect the other.

Choose a guide

Task Guide
Send one text request through the shared provider client LLM client
Find and run a local ONNX model Local models
Use Phi Silica through Windows AI Windows AI
Check which configured models respond Model probing
Inspect provider limits and rank probe results Provider limits
Compare models or run coding benchmarks Benchmarks

Package map

Path Purpose
tools/llm/llm_client.py Static text-generation facade
tools/llm/provider_adapters.py Provider-specific calls
tools/llm/model_probe.py Availability checks and probe cache
tools/llm/probe_discovery.py Provider discovery
tools/llm/model_inventory.py Inventory assembly
tools/llm/local_model.py ONNX Runtime GenAI wrapper
tools/llm/windows_ai.py Windows AI bridge wrapper
tools/llm/model_bakeoff.py Small cross-model comparison
tools/llm/check_provider_limits.py Provider metadata checks
tools/llm/rank_models.py Ranking from probe and limit files
tools/agents/benchmarks/ Benchmark definitions, loaders, runners, and evaluation helpers
tools/core/ Configuration, errors, response cache, prompt storage, and utilities

Install

From the repository root:

python -m pip install -e .

The root package requires Python 3.11 or newer.

Optional features have additional requirements:

  • local ONNX: install the appropriate onnxruntime-genai package for the target hardware;
  • provider-limit inspection: install requests;
  • remote provider SDKs or local runtimes: follow the provider-specific guide.

The repository-wide setup command installs all local packages:

just setup

The current justfile uses PowerShell recipes. On Linux or macOS, follow the manual commands in Installation.

Keep results reproducible

When using discovery, probes, rankings, or benchmarks, record:

  • the exact model identifier;
  • provider endpoint and model revision when available;
  • prompt or task revision;
  • temperature and output limit;
  • probe and benchmark timestamp;
  • whether remote providers were enabled; and
  • the generated JSON result file.

Availability and provider limits change. A saved probe is evidence for its timestamp, not a permanent model catalog.