ContextAtlas

Stable, reusable, and observable code context infrastructure for AI agents

Hybrid Retrieval · Project Memory · MCP Server · Retrieval Observability

简体中文 · Docs · First Use · Deployment · CLI · MCP

ContextAtlas is an open-source context infrastructure for AI coding agents — providing hybrid code retrieval, project memory, and retrieval observability as a CLI, MCP server, or embeddable library. It combines tree-sitter semantic chunking, LanceDB vector search, SQLite FTS5 full-text search, and token-aware context packing to deliver structured, high-quality code context to tools like Claude Code, Codex, and custom agent workflows.

Updates

• 2026-04-06: tightened the default user path, memory governance, and operational visibility to make first use, feedback loops, and health checks clearer.
• 2026-04-07: improved the indexing pipeline with lighter planning, snapshot copy reduction, queue observability, fallback hardening, and repeatable benchmarks.
• 2026-04-08: added the embedding gateway, local caching and multi-upstream routing, plus Hugging Face integration and MCP context lifecycle tools.
• 2026-04-09: added churn / cost-aware index planning, moved long-term memory into dedicated tables + FTS5, and finished default-path hardening, threshold configuration, ops alert threshold alignment, and doc sync.

Contents

• Why ContextAtlas
• Where it fits
• Core Capabilities
• Tech Stack
• Installation
• Configuration
• Quick Start
• Integration Modes
• Architecture Overview
• Documentation
• Friendly Links
• License

ContextAtlas is not just a code search tool. It addresses a more practical engineering problem:

• can an agent find the right code faster in a large repository?
• can repository understanding be persisted instead of rediscovered every session?
• can retrieval, indexing, and memory quality be observed and improved over time?

If you are building Claude Code workflows, MCP clients, or custom agent systems, ContextAtlas provides a context infrastructure layer: retrieval, memory, context packing, and observability.

Why ContextAtlas

In real projects, agent failures are often not caused by a weak model. They come from weak context systems:

• the relevant code is not found
• the returned code is too fragmented and lacks surrounding context
• the same module has to be re-understood again and again
• indexes become stale, retrieval quality degrades, and token budgets get exhausted without clear signals

ContextAtlas turns this into a composable set of capabilities:

• Find: hybrid retrieval narrows down the relevant implementation
• Expand: graph expansion and token packing turn hits into usable local context
• Store: project memory, long-term memory, and a cross-project hub preserve knowledge
• Observe: health checks, telemetry, usage analysis, and alerts make the system diagnosable

Where it fits

• As a repository retrieval backend for coding agents
• As an MCP server for external clients that need code retrieval and memory tools
• As a local CLI / skill backend for scripts, CI, and workflow automation
• As a cross-project knowledge layer for reusable module knowledge and decision history

Core capabilities

ContextAtlas decides what context to provide, not how the task should be executed. It does not handle agent reasoning, workflow orchestration, or business API actions.

Tech stack

• TypeScript / Node.js 20+
• Tree-sitter for semantic chunking
• SQLite + FTS5 for metadata, retrieval, queues, and memory hub storage
• LanceDB for vector storage
• Model Context Protocol SDK for MCP integration

Installation

npm install -g @codefromkarl/context-atlas

Product identity mapping:

• Repository: ContextAtlas
• npm package: @codefromkarl/context-atlas
• CLI command: contextatlas

Available commands:

• contextatlas
• cw (short alias)

The docs use contextatlas as the primary command name. cw remains as a compatibility alias.

Configuration

Initialize the config directory and example environment file first:

contextatlas init

Default config file location:

~/.contextatlas/.env

Minimum required configuration:

EMBEDDINGS_API_KEY=
EMBEDDINGS_BASE_URL=
EMBEDDINGS_MODEL=

RERANK_API_KEY=
RERANK_BASE_URL=
RERANK_MODEL=

Index update planning also supports these optional knobs:

INDEX_UPDATE_CHURN_THRESHOLD=0.35
INDEX_UPDATE_COST_RATIO_THRESHOLD=0.65
INDEX_UPDATE_MIN_FILES=8
INDEX_UPDATE_MIN_CHANGED_FILES=5

• INDEX_UPDATE_CHURN_THRESHOLD: when the changed-file ratio crosses this value, index:plan / index:update will favor full
• INDEX_UPDATE_COST_RATIO_THRESHOLD: triggers full when the estimated incremental cost is close to a full rebuild
• INDEX_UPDATE_MIN_FILES / INDEX_UPDATE_MIN_CHANGED_FILES: require both repo size and change size to clear a minimum bar before escalation is allowed

init writes an editable example .env, including default SiliconFlow endpoints and recommended model settings.

Quick start

If you are onboarding for the first time, start with the First use guide.

1) Confirm the default entry flow

contextatlas start /path/to/repo

2) Initialize and fill in API settings

contextatlas init
# edit ~/.contextatlas/.env

3) Index a repository

contextatlas index /path/to/repo

4) Run local retrieval

contextatlas search \
  --repo-path /path/to/repo \
  --information-request "How is the authentication flow implemented?"

5) Start the daemon (recommended)

contextatlas daemon start

6) Expose it as an MCP server

contextatlas mcp

Integration modes

1. As a local CLI / skill backend

Useful for:

• custom agent skills
• shell workflows and CI scripts
• local debugging and retrieval analysis

Example:

# retrieval
contextatlas search --repo-path /path/to/repo --information-request "Where is the payment retry policy implemented?"

# project memory
contextatlas memory:find "search"
contextatlas decision:list

# health
contextatlas health:full

2. As an MCP server

Useful for:

• desktop clients that support MCP
• agent systems that need standard tool-based access to ContextAtlas capabilities

Claude Desktop configuration example:

{
  "mcpServers": {
    "contextatlas": {
      "command": "contextatlas",
      "args": ["mcp"]
    }
  }
}

ContextAtlas MCP tools cover:

• code retrieval
• project memory
• long-term memory
• cross-project hub operations
• auto-recording and memory suggestion flows

Architecture overview

Indexing:  Crawler / Scanner → Chunking → Indexing → Vector / SQLite Storage
Retrieval: Vector + FTS Recall → RRF → Rerank → Graph Expansion → Context Packing
Memory:    Project Memory / Long-term Memory / Hub → CLI / MCP Tools

ContextAtlas focuses on what context to provide, not how the task should be executed. For a fuller architecture explanation, see repository positioning and engineering positioning.

Documentation map

Contributing

Ways to improve ContextAtlas:

• open issues for bugs or documentation gaps
• submit PRs for retrieval, memory, monitoring, or documentation improvements
• contribute real-world usage patterns, deployment notes, and benchmark data
• improve README, CLI docs, and MCP examples

Before submitting code, it helps to:

1. run pnpm build and make sure the repo still builds
2. keep command examples, README, and docs aligned with the implementation
3. update functionality, documentation, and operational notes together when possible

Development

pnpm build
pnpm build:release
pnpm dev
node dist/index.js

Friendly links

https://linux.do/

License

MIT