How Driver Works
Driver processes your codebase before you ask a question. Answers are fast, complete, and grounded in your code.
At the core is the Transpiler: an ahead-of-time compiler that transforms source code into structured, queryable context. Unlike runtime search tools that explore code on the fly, Driver pre-computes a complete understanding of your codebase and keeps it current as code changes.
The approach
The Transpiler is built on four principles:
Systematic structural decomposition
Driver decomposes code into file tree DAGs and syntax trees upfront. Every file, directory, function, class, and interface is mapped into a structured graph before any analysis begins. This ensures nothing is missed and relationships between components are preserved.
Exhaustive multi-pass processing
Multiple passes at every level of abstraction: symbol-level, refined, and synthesized. The first pass extracts raw structure. Subsequent passes build higher-level understanding, from individual function documentation to architecture-wide patterns. Each pass builds on the previous one.
Pre-computed content for downstream use
Architecture docs, code maps, and guides are built ahead of time for AI agent consumption. When your agent asks about a codebase, the answer already exists. There is no waiting for analysis to run, no token budget spent on discovery.
High signal-to-noise ratio distillation
Keeping irrelevant content out is as important as getting relevant content in. Driver's multi-pass processing distills raw code into dense, high-signal documentation. The result: agents get exactly what they need without noise that causes hallucinations.
What gets produced
For every connected codebase, the Transpiler generates:
- Architecture overviews: system structure, component relationships, key design decisions
- Code maps: every file and directory annotated with descriptions
- File-level documentation: functions, classes, interfaces, with symbol-level detail
- Changelogs: derived from git history, organized by month
- Onboarding guides: optimized for AI agents to get oriented fast
These artifacts update automatically when your code changes. Version control integrations (GitHub, GitLab, Bitbucket, Azure DevOps) trigger incremental updates, so only changed files are reprocessed.
Product architecture
Driver's infrastructure has three layers:
Context Engine (core): Processes source code into structured, complete understanding. Produces architecture docs, code maps, changelogs, and symbol-level documentation.
Governance Control Plane: Controls who sees what. SCM integration, identity and access management (IdP, SCIM, RBAC), and audit logging for compliance.
Application Integrations: Delivers context where you work. MCP tools for AI coding assistants, APIs for CI/CD automation, and the Driver web application for browsing and exploration.