Skip to content
context-engine

Introduction

The semantic interface between coding agents and live codebases.

Context Engine gives coding agents a semantic interface to the live codebase. Instead of reconstructing relationships from raw files, an agent can navigate symbols, resolve types and consumers, and fetch only the source ranges the current task needs.

Files store text. Agents need code.

A file is the storage format, but code has structure that does not live at fixed line coordinates:

  • symbols have exact boundaries and meaningful parts;
  • definitions connect to callers, implementations, and dependencies;
  • types and parameter names form contracts;
  • a changing workspace must remain synchronized with its language servers.

Context Engine makes those relationships directly addressable. Structural analysis maps the codebase, language servers resolve semantic identity, and handles let the agent request a specific symbol again without repeating discovery.

The agent chooses what to fetch

Context Engine does not rank text chunks and silently discard the rest. The agent asks an explicit question—such as “what does this file expose?”, “where is this identifier defined?”, or “which code uses this symbol?”—and receives the corresponding source-backed context.

That keeps selection under the agent’s control:

Starting pointFirst tool
A diagnostic, stack trace, or exact file:lineline_context
An identifier visible in sourcejump
A symbol name with no known locationgrep_definition
A known file whose structure is unfamiliaroutline
A handle from an earlier resultextract
A symbol whose consumers mattershow_usage
A visible identifier whose type or documentation is enoughhover
A known exact line rangesmart_read

The tools are designed to compose. A typical investigation moves from discovery to a precise symbol, then follows only the definitions, usages, or implementation details needed to complete the task.

Live, local semantics

Analysis runs locally against the project files and configured language servers. Context Engine combines a polyglot structural layer with live language-server semantics, then watches the workspace so later requests reflect current source rather than a pre-built semantic snapshot.

Dependencies participate in the same navigation model as workspace code. An agent can follow a visible symbol into the standard library or a third-party package and inspect the API or implementation selected by the project.

Continue

Follow the Quickstart to connect Context Engine to a coding client, or open a tool page from the sidebar when you already know the operation you need.