Codebase intelligence.
Beyond one repository.

One MCP server for codebase intelligence across agents, IDEs, and any model. Connect your repositories and organization knowledge so every workflow starts with the right context. Bring your own factory; keep your intelligence connected.

Get started Plans from $5/month

View on GitHub NPM Bridge

MCP One open protocol

Any model Through your agent or IDE

Cross-repo Organization-wide context

MCP · cross_repo_search

// Your agent asks across the business
cross_repo_search({
  "query": "Who consumes CheckoutCreated?",
  "target_repos": ["storefront", "billing", "shared"],
  "trace_boundary": true
})

// Illustrative source-backed context
shared/events.ts       → event contract
storefront/checkout.ts → publisher
billing/invoices.ts    → consumer

Self-Learning Adapts from feedback when enabled

Shared Knowledge Context that travels across tools

Measured with real agent tasks · internal evaluation · October 20, 2026

Fewer calls. Fewer turns. Lower cost.

26% fewertool calls per task, on average

25% feweragent turns per task, on average

20% lowermodel cost per task, on average

About the evaluation

Claude Sonnet 5.5 with and without Context Engine: 80 paired runs (40 code-navigation tasks, two seeds) in two repositories. Tool calls per task fell from 14.2 to 10.4 and agent turns from 15.2 to 11.4 (both p < 0.001); model cost per task fell from $0.118 to $0.094 (p = 0.02). Answer accuracy held level (60.5% vs 61.3%). Intervals are task-clustered bootstrap; p-values are task-level sign-flip permutation tests. An earlier October 5, 2026 evaluation with Claude Sonnet 4.5 measured file-and-symbol accuracy rising from 52.2% to 64.3%.

With learning enabled, retrieval adapts as feedback accumulates. These measurements capture the tested model, configuration and sessions.

Public retrieval-benchmark references: CoIR code-retrieval benchmark ↗, our benchmark methodology (January 2026) ↗, and TokensToGold methodology and scoring code ↗. The Context Engine figures above come from our separate agent A/B evaluation.

How It Works

Code, documentation, relationships and memory become grounded context your AI tools can use. Connect through MCP without tying your business knowledge to a single model.

Raw Sources Semantic Processing MCP Tools

Code Files

Dependencies

Documentation

Git History

Team Knowledge

Helion Index

Embeddings

Symbol Graph

Memory Store

repo_search

symbol_graph

context_answer

memory_find

Intelligence for your whole codebase

From one repository to knowledge across your business. Keep context portable as your agents, IDEs and models evolve.

Semantic Search

Find the implementation behind an idea, not just a matching keyword. Give your agent relevant source evidence instead of an entire repository.

Symbol Graph

Navigate definitions, callers, callees and dependencies. Understand what a change touches before your agent writes it.

Cross-Repo Intelligence

Connect repositories across your business. Trace APIs, shared types and dependencies from the frontend through services and libraries.

Organization Knowledge

Build a shared ontology: a map of services, concepts, relationships and decisions grounded in your code and documentation. Keep useful team knowledge in persistent memory.

Any Agent. Any IDE. Any Model.

One MCP connection for Claude, Codex, Cursor, Copilot and other compatible clients. Choose your model through your agent or IDE; keep your context portable.

Your Knowledge, Your Control

Workspace permissions keep organization knowledge scoped to the right people. Choose hosted access or self-host on your own infrastructure.

Your workflow. One MCP connection.

Bring Context Engine to your agents and editors. Your organization’s context stays connected when your tools change.

VS Code Marketplace ↗ Open VSX ↗ MCP bridge on npm ↗ MCP setup & documentation ↗

BYOF — Bring Your Own Factory

Your agents run the factory. Context Engine supplies shared intelligence.

Give planning, coding, review and verification agents a common view of your repositories, dependencies and team knowledge through MCP-compatible runtimes. Choose your orchestration layer instead of rebuilding your context for every agent.

Explore SWE-AF, an independent AgentField software factory ↗

Start with $5/month

Every hosted plan requires a subscription. Choose the workspace size that fits your team.