Skip to main content
CruciHiL includes AI in three places:

analyze_component

Extract the signal interface contract of a C/C++ firmware SWC. Maps AUTOSAR RTE and COM module identifiers to DBC signal names.

crucihil discover

Probe hardware and generate a ready-to-use rig TOML. Describes CAN interfaces, bitrates, and ECU topology automatically.

generate_test_suite

Scaffold a YAML manifest and matching Python test stubs from a natural-language description, with signal names auto-extracted from your DBC.

​
BYOK — Bring Your Own Key

CruciHiL does not include or host an AI model. Every AI feature requires an API key from a supported provider. This means:
  • You control the model and the cost
  • No data leaves your environment via CruciHiL — queries go directly from your machine to the AI provider
  • You can use whichever model your team already has access to

​
Supported providers

ProviderEnvironment variableDefault model
Anthropic (Claude)ANTHROPIC_API_KEYclaude-sonnet-4-6
OpenAI (GPT)OPENAI_API_KEYgpt-4o
Google (Gemini)GOOGLE_API_KEYgemini-2.0-flash

​
Provider auto-detection

All AI features check for API keys in this order:
  1. ANTHROPIC_API_KEY
  2. OPENAI_API_KEY
  3. GOOGLE_API_KEY
The first key found determines the provider. No configuration needed beyond setting the environment variable. Override with --provider if you want to force a specific provider: 
crucihil analyze --source swc/brake --component Brake \
  --rig rigs/bench.toml --provider openai

​
Setting up API keys

export ANTHROPIC_API_KEY=sk-ant-api03-...
Add to your shell profile (~/.bashrc, ~/.zshrc) to persist across sessions.

​
No AI key? No problem

All AI features degrade gracefully when no key is present:
  • crucihil analyze — returns the extracted identifiers and corpus size without signal matching. You can review identifiers manually.
  • crucihil discover — falls back to --no-ai mode and generates a minimal stub TOML with # TODO: markers.
  • generate_test_suite — writes static placeholder stubs instead of AI-generated assertions.

​
MCP + AI = stronger together

The MCP server exposes 14 tools to AI assistants (Claude Desktop, Copilot, Cursor). When an AI assistant has access to both the MCP tools and a real CruciHiL deployment, it can:
  • Query which rigs are connected
  • Trigger test runs and poll for results
  • Explain why a specific test failed, with signal trace context
  • Generate new test suites from natural language
  • Map firmware SWC source to DBC signals
This is the primary workflow for AI-assisted firmware testing — the AI orchestrates, CruciHiL executes.

​
What the AI sees

CruciHiL sends the AI:
  • For analyze_component: a filtered list of C/C++ identifiers (up to 500) and the DBC signal corpus (all MessageName.SignalName pairs). No source code text.
  • For discover: the output of hardware probe commands (ip link, /proc/net/dev) plus any description you provide.
  • For generate_test_suite: signal names from your DBC, hardware capability context from the rig TOML, and your natural-language description.
Source code is never sent to the AI. Only identifiers extracted by tree-sitter are included. This is a deliberate design decision to keep sensitive firmware IP within your environment.

​
See also