Context Engine MCP Server

A local-first, agent-agnostic Model Context Protocol (MCP) server implementation using the Auggie SDK as the core context engine.

📚 New here? Check out INDEX.md for a complete documentation guide!

🚀 Quick Start: QUICKSTART.md → GETTING_STARTED.md → API_REFERENCE.md

🏗️ Architecture: TECHNICAL_ARCHITECTURE.md for deep technical dive

Architecture

This implementation follows a clean 5-layer architecture as outlined in plan.md:

┌────────────────────────────┐
│ Coding Agents (Clients)    │  Layer 4: Claude, Cursor, etc.
│ Codex | Claude | Cursor    │
└────────────▲───────────────┘
             │ MCP (tools)
┌────────────┴───────────────┐
│ MCP Interface Layer        │  Layer 3: server.ts, tools/
│ (standardized tool API)    │
└────────────▲───────────────┘
             │ internal API
┌────────────┴───────────────┐
│ Context Service Layer      │  Layer 2: serviceClient.ts
│ (query orchestration)      │
└────────────▲───────────────┘
             │ domain calls
┌────────────┴───────────────┐
│ Core Context Engine        │  Layer 1: Auggie SDK
│ (indexing, retrieval)      │
└────────────▲───────────────┘
             │ storage
┌────────────┴───────────────┐
│ Storage / Index Backend    │  Layer 5: Auggie's internal
│ (vectors, metadata)        │
└────────────────────────────┘

Layer Responsibilities

• Layer 1 (Core Engine): Auggie SDK handles file ingestion, chunking, embedding, and semantic retrieval
• Layer 2 (Service): Orchestrates context, formats snippets, deduplicates, enforces limits
• Layer 3 (MCP Interface): Exposes tools, validates I/O, maps calls to service layer
• Layer 4 (Agents): Consume context and generate responses
• Layer 5 (Storage): Persists embeddings and metadata

Features

MCP Tools (41 tools available)

Core Context Tools (10)

1. index_workspace(force?) - Index workspace files for semantic search
   • force (optional): Force re-indexing even if files haven't changed
2. codebase_retrieval(query, top_k?) - PRIMARY semantic search with JSON output for programmatic use
   • query: Natural language search query
   • top_k (optional): Number of results to return (default: 5)
3. semantic_search(query, top_k?, mode?, bypass_cache?, timeout_ms?) - Semantic code search with markdown-formatted output
   • query: Natural language search query
   • top_k (optional): Number of results to return (default: 5)
   • mode (optional): "fast" (default) or "deep" for higher recall at higher latency
   • bypass_cache (optional): When true, bypass caches for this call
   • timeout_ms (optional): Cap time spent in retrieval pipeline (ms)
4. get_file(path) - Retrieve complete file contents
   • path: Relative path to file from workspace root
5. get_context_for_prompt(query, max_files?, token_budget?, include_related?, min_relevance?, bypass_cache?) - Get comprehensive context bundle for prompt enhancement
   • query: Context request description
   • max_files (optional): Maximum files to include (default: 5)
   • token_budget (optional): Token budget for the bundle (default: 8000)
   • include_related (optional): Include related/imported files (default: true)
   • min_relevance (optional): Minimum relevance score (default: 0.3)
   • bypass_cache (optional): When true, bypass caches for this call
6. enhance_prompt(prompt) - AI-powered prompt enhancement with codebase context
   • prompt: Simple prompt to enhance
7. index_status() - View index health metadata (status, fileCount, lastIndexed, isStale)
8. reindex_workspace() - Clear and rebuild the entire index from scratch
9. clear_index() - Remove index state without rebuilding
10. tool_manifest() - Discovery tool for available capabilities

Memory System (2)

11. add_memory(category, content, title?) - Store persistent memories for future sessions
    • category: 'preferences', 'decisions', or 'facts'
    • content: The memory content to store (max 5000 characters)
    • title (optional): Title for the memory
12. list_memories(category?) - List all stored memories
    • category (optional): Filter to a specific category

Planning & Execution (4)

13. create_plan(task, options?) - Generate structured execution plans with DAG analysis
    • task: Task or goal to plan for
    • generate_diagrams (optional): Generate Mermaid diagrams (default: true)
14. refine_plan(current_plan, feedback?, clarifications?) - Refine existing plans based on feedback
15. visualize_plan(plan, diagram_type?) - Generate visual representations (Mermaid diagrams)
16. execute_plan(plan, ...) - Execute plan steps with AI-powered code generation

Plan Management (13)

17. save_plan(plan, name?, tags?, overwrite?) - Save plans to persistent storage
18. load_plan(plan_id \| name) - Load previously saved plans
19. list_plans(status?, tags?, limit?) - List saved plans with filtering
20. delete_plan(plan_id) - Delete saved plans from storage
21. request_approval(plan_id, step_numbers?) - Create approval requests for plans or specific steps
22. respond_approval(request_id, action, comments?) - Respond to approval requests
23. start_step(plan_id, step_number) - Mark a step as in-progress
24. complete_step(plan_id, step_number, notes?, files_modified?) - Mark a step as completed
25. fail_step(plan_id, step_number, error, ...) - Mark a step as failed
26. view_progress(plan_id) - View execution progress and statistics
27. view_history(plan_id, limit?, include_plans?) - View version history of a plan
28. compare_plan_versions(plan_id, from_version, to_version) - Generate diff between versions
29. rollback_plan(plan_id, version, reason?) - Rollback to a previous plan version

Code Review (5)

30. review_changes(diff, file_contexts?, options?) - AI-powered code review with structured output
31. review_git_diff(target?, base?, include_patterns?, options?) - Review code changes from git automatically
32. review_diff(diff, changed_files?, options?) - Enterprise review with risk scoring and static analysis
    • Risk scoring (1-5) based on deterministic preflight
    • Change classification (feature/bugfix/refactor/infra/docs)
    • Optional static analysis (TypeScript, Semgrep)
    • Per-phase timing telemetry
33. check_invariants(diff, changed_files?, invariants_path?) - Run YAML invariants deterministically (no LLM)
34. run_static_analysis(changed_files?, options?) - Run local static analyzers (tsc, semgrep)

Reactive Review (7)

35. reactive_review_pr(...) - Start a session-based, parallelized code review
36. get_review_status(session_id) - Track progress of a reactive review
37. pause_review(session_id) - Pause a running review session
38. resume_review(session_id) - Resume a paused session
39. get_review_telemetry(session_id) - Detailed metrics (tokens, speed, cache hits)
40. scrub_secrets(content) - Mask API keys and sensitive data
41. validate_content(content, content_type, ...) - Multi-tier validation for AI-generated content

Key Characteristics

• ✅ Local-first: No cloud dependencies, no exposed ports, no data leakage
• ✅ Agent-agnostic: Works with any MCP-compatible coding agent
• ✅ LLM-agnostic: No LLM-specific logic in the engine
• ✅ Storage-agnostic: Auggie SDK handles storage abstraction
• ✅ Extensible: Clean separation allows easy feature additions
• ✅ Real-time watching: Automatic incremental indexing on file changes (v1.1.0)
• ✅ Background indexing: Non-blocking indexing via worker threads (v1.1.0)
• ✅ Offline policy: Enforce local-only operation with environment variable (v1.1.0)
• ✅ Planning mode: AI-powered implementation planning with DAG analysis (v1.4.0)
• ✅ Execution tracking: Step-by-step execution with dependency management (v1.4.0)
• ✅ Version control: Plan versioning with diff and rollback support (v1.4.0)
• ✅ Approval workflows: Built-in approval system for plans and steps (v1.4.0)
• ✅ Defensive programming: Comprehensive null/undefined handling (v1.4.1)
• ✅ Cross-session memory: Persistent memory system for preferences, decisions, and facts (v1.4.1)
• ✅ AI-powered code review: Structured code review with confidence scoring and priority levels (v1.7.0)
• ✅ Git integration: Automatic diff retrieval for staged, unstaged, branch, and commit changes (v1.7.0)
• ✅ Reactive Optimization: 180-600x faster reactive reviews via AI Agent Executor, Multi-layer Caching, Batching, and Worker Pool Optimization (v1.8.0)
• ✅ High Availability: Circuit breakers, adaptive timeouts, and zombie session detection (v1.8.0)
• ✅ Static analysis integration: Optional TypeScript and Semgrep analyzers for deterministic feedback (v1.9.0)
• ✅ Invariants checking: YAML-based custom rules for deterministic code review (v1.9.0)
• ✅ Per-phase telemetry: Detailed timing breakdowns for review pipeline optimization (v1.9.0)

Reactive Review Optimizations (v1.8.0)

Version 1.8.0 introduces massive performance improvements to the reactive code review system, reducing review times from 30-50 minutes to 3-15 seconds for typical PRs.

Optimization Stack

Phase	Feature	Performance Gain	Description
Phase 1	AI Agent Executor	15-50x	Executes reviews directly via the AI agent instead of external API calls.
Phase 2	Multi-Layer Cache	2-4x (cached)	3-layer system: Memory (fastest) -> Commit (git-aware) -> File Hash (content-based).
Phase 3	Continuous Batching	2-3x	Accumulates and processes multiple files in a single AI request.
Phase 4	Worker Pool Optimization	1.5-2x	CPU-aware parallel execution with intelligent load balancing.

Total Performance Improvement

Scenario	v1.7.1	v1.8.0	Improvement
Cold Run (10 steps)	30-50 min	~60-90 sec	25-45x ⚡
Cached Run	30-50 min	~10-30 sec	60-180x ⚡
Batched Run	30-50 min	~5-15 sec	120-360x ⚡
Full Optimization	30-50 min	3-10 sec	180-600x 🚀

Static Analysis & Invariants (v1.9.0)

Version 1.9.0 introduces optional static analysis and deterministic invariants checking for enhanced code review capabilities.

Static Analysis Features

Analyzer	Description	Opt-in
TypeScript	Type checking via tsc --noEmit	Default
Semgrep	Pattern-based security/quality checks	Optional (requires installation)

Usage

Enable Static Analysis in review_diff

review_diff({
  diff: "<unified diff>",
  changed_files: ["src/file.ts"],
  options: {
    enable_static_analysis: true,
    static_analyzers: ["tsc", "semgrep"],
    static_analysis_timeout_ms: 60000
  }
})

Run Static Analysis Standalone

run_static_analysis({
  changed_files: ["src/file.ts"],
  options: {
    analyzers: ["tsc", "semgrep"],
    timeout_ms: 60000,
    max_findings_per_analyzer: 20
  }
})

Check Custom Invariants

check_invariants({
  diff: "<unified diff>",
  changed_files: ["src/file.ts"],
  invariants_path: ".review-invariants.yml"
})

Invariants Configuration

Create .review-invariants.yml in your workspace root:

invariants:
  - id: no-console-log
    pattern: "console\\.log"
    message: "Remove console.log statements before committing"
    severity: MEDIUM

  - id: no-todo-comments
    pattern: "TODO|FIXME"
    message: "Resolve TODO/FIXME comments"
    severity: LOW

  - id: require-error-handling
    pattern: "catch\\s*\\(\\s*\\)"
    message: "Empty catch blocks should log or handle errors"
    severity: HIGH

Benefits

• ✅ Deterministic: No LLM required for invariants/static analysis
• ✅ Fast: Local execution, no API calls
• ✅ CI-Friendly: Structured JSON output suitable for CI/CD pipelines
• ✅ Customizable: YAML-based rules, configurable analyzers
• ✅ Opt-in: Disabled by default, enable as needed

Per-Phase Telemetry

The review_diff tool now reports detailed timing breakdowns in stats.timings_ms:

{
  "stats": {
    "timings_ms": {
      "preflight": 45,
      "invariants": 12,
      "static_analysis": 3200,
      "context_fetch": 890,
      "secrets_scrub": 5,
      "llm_structural": 1200,
      "llm_detailed": 2400
    }
  }
}

This allows you to:

• Identify performance bottlenecks in the review pipeline
• Optimize timeout settings for your workflow
• Monitor static analysis overhead
• Track LLM usage patterns

Planning Workflow (v1.4.0+)

The Context Engine now includes a complete planning and execution system:

1. Create a Plan

create_plan({
  task: "Implement user authentication with JWT tokens",
  generate_diagrams: true
})

2. Save the Plan

save_plan({
  plan: "<plan JSON>",
  name: "JWT Authentication",
  tags: ["auth", "security"]
})

3. Execute Step-by-Step

// Start a step
start_step({ plan_id: "plan_abc123", step_number: 1 })

// Complete it
complete_step({
  plan_id: "plan_abc123",
  step_number: 1,
  notes: "Created User model"
})

// Check progress
view_progress({ plan_id: "plan_abc123" })

4. Track History

// View version history
view_history({ plan_id: "plan_abc123" })

// Compare versions
compare_plan_versions({
  plan_id: "plan_abc123",
  from_version: 1,
  to_version: 2
})

// Rollback if needed
rollback_plan({ plan_id: "plan_abc123", version: 1 })

See EXAMPLES.md for complete planning workflow examples.

Memory System (v1.4.1)

The Context Engine includes a cross-session memory system that persists preferences, decisions, and project facts across sessions.

Memory Categories

Category	Purpose	Examples
preferences	Coding style and tool preferences	"Prefer TypeScript strict mode", "Use Jest for testing"
decisions	Architecture and design decisions	"Chose JWT over sessions", "Using PostgreSQL"
facts	Project facts and environment info	"API runs on port 3000", "Uses monorepo structure"

Adding Memories

// Store a preference
add_memory({
  category: "preferences",
  content: "Prefers functional programming patterns over OOP"
})

// Store an architecture decision with a title
add_memory({
  category: "decisions",
  title: "Authentication Strategy",
  content: "Chose JWT with refresh tokens for stateless authentication. Sessions were considered but rejected due to horizontal scaling requirements."
})

// Store a project fact
add_memory({
  category: "facts",
  content: "The API uses PostgreSQL 15 with pgvector extension for embeddings"
})

Automatic Memory Retrieval

Memories are automatically included in get_context_for_prompt results when relevant:

// Memories are retrieved alongside code context
const context = await get_context_for_prompt({
  query: "How should I implement authentication?"
})
// Returns: code context + relevant memories about auth decisions

Memory Files

Memories are stored in .memories/ as markdown files:

• preferences.md - Coding style preferences
• decisions.md - Architecture decisions
• facts.md - Project facts

These files are human-editable and can be version controlled with Git.

Prerequisites

1. Node.js 18+
2. Auggie CLI - Install globally:

   npm install -g @augmentcode/auggie
   

3. Authentication - Run auggie login or set environment variables:

   export AUGMENT_API_TOKEN="your-token"
   export AUGMENT_API_URL="https://api.augmentcode.com"
   

Installation

# Clone or navigate to the repository
cd context-engine

# Install dependencies
npm install

# Build the project
npm run build

Usage

Standalone Mode

Using the Management Script (Windows)

For Windows users, a convenient batch file is provided for managing the server:

# Start the server with indexing and file watching
manage-server.bat start

# Check server status
manage-server.bat status

# Restart the server
manage-server.bat restart

# Stop the server
manage-server.bat stop

The management script automatically:

• Uses the current directory as workspace
• Enables indexing (--index)
• Enables file watching (--watch)
• Logs output to .server.log
• Tracks the process ID in .server.pid

Manual Start (All Platforms)

# Start server with current directory
node dist/index.js

# Start with specific workspace
node dist/index.js --workspace /path/to/project

# Index workspace before starting
node dist/index.js --workspace /path/to/project --index

# Enable file watcher for automatic incremental indexing (v1.1.0)
node dist/index.js --workspace /path/to/project --watch

CLI Options

Option	Alias	Description
--workspace <path>	-w	Workspace directory to index (default: current directory)
--index	-i	Index the workspace before starting server
--watch	-W	Enable filesystem watcher for incremental indexing
--http	-	Enable HTTP server (in addition to stdio)
--http-only	-	Enable HTTP server only (for VS Code integration)
--port <port>	-p	HTTP server port (default: 3333)
--help	-h	Show help message

With Codex CLI

1. Build the project:

   npm run build
   

2. Add the MCP server to Codex CLI:

   codex mcp add context-engine -- node /absolute/path/to/context-engine/dist/index.js --workspace /path/to/your/project
   

   Or edit ~/.codex/config.toml directly:

   [mcp_servers.context-engine]
   command = "node"
   args = [
       "/absolute/path/to/context-engine/dist/index.js",
       "--workspace",
       "/path/to/your/project"
   ]
   

3. Restart Codex CLI

4. Type /mcp in the TUI to verify the server is connected

With Other MCP Clients (Antigravity, Claude Desktop, Cursor)

For other MCP clients, add this server to your client's MCP configuration:

{
  "mcpServers": {
    "context-engine": {
      "command": "node",
      "args": [
        "/absolute/path/to/context-engine/dist/index.js",
        "--workspace",
        "/path/to/your/project"
      ]
    }
  }
}

See QUICKSTART.md - Step 5B for detailed instructions for each client.

Development

# Watch mode for development
npm run dev

# Build for production
npm run build

# Run the server
npm start

Project Structure

context-engine/
├── src/
│   ├── index.ts              # Entry point with CLI parsing
│   ├── mcp/
│   │   ├── server.ts         # MCP server implementation
│   │   ├── serviceClient.ts  # Context service layer
│   │   ├── tools/
│   │   │   ├── index.ts      # index_workspace tool
│   │   │   ├── search.ts     # semantic_search tool
│   │   │   ├── file.ts       # get_file tool
│   │   │   ├── context.ts    # get_context_for_prompt tool
│   │   │   ├── enhance.ts    # enhance_prompt tool
│   │   │   ├── status.ts     # index_status tool (v1.1.0)
│   │   │   ├── lifecycle.ts  # reindex/clear tools (v1.1.0)
│   │   │   ├── manifest.ts   # tool_manifest tool (v1.1.0)
│   │   │   ├── plan.ts       # Planning tools (v1.4.0)
│   │   │   └── planManagement.ts  # Plan persistence/workflow tools (v1.4.0)
│   │   ├── services/         # Business logic services (v1.4.0)
│   │   │   ├── planningService.ts        # Plan generation, DAG analysis
│   │   │   ├── planPersistenceService.ts # Save/load/list plans
│   │   │   ├── approvalWorkflowService.ts # Approval request handling
│   │   │   ├── executionTrackingService.ts # Step progress tracking
│   │   │   └── planHistoryService.ts     # Version history, rollback
│   │   ├── types/            # TypeScript type definitions (v1.4.0)
│   │   │   └── planning.ts   # Planning-related types
│   │   └── prompts/          # AI prompt templates (v1.4.0)
│   │       └── planning.ts   # Planning system prompts
│   ├── watcher/              # File watching (v1.1.0)
│   │   ├── FileWatcher.ts    # Core watcher logic
│   │   ├── types.ts          # Event types
│   │   └── index.ts          # Exports
│   └── worker/               # Background indexing (v1.1.0)
│       ├── IndexWorker.ts    # Worker thread
│       └── messages.ts       # IPC messages
├── tests/                    # Unit tests (186 tests)
├── plan.md                   # Architecture documentation
├── package.json
├── tsconfig.json
└── README.md

Example Usage

Once connected to Codex CLI, you can use natural language:

• "Search for authentication logic in the codebase"
• "Show me the database schema files"
• "Get context about the API endpoints"
• "Find error handling patterns"

The server will automatically use the appropriate tools to provide relevant context.

Environment Variables

Variable	Description	Default
AUGMENT_API_TOKEN	Auggie API token (or use auggie login)	-
AUGMENT_API_URL	Auggie API URL	https://api.augmentcode.com
CONTEXT_ENGINE_OFFLINE_ONLY	Enforce offline-only policy (v1.1.0)	false
REACTIVE_ENABLED	Enable reactive review features	false
REACTIVE_USE_AI_AGENT_EXECUTOR	Use local AI agent for reviews (Phase 1)	false
REACTIVE_ENABLE_MULTILAYER_CACHE	Enable 3-layer caching (Phase 2)	false
REACTIVE_ENABLE_BATCHING	Enable request batching (Phase 3)	false
REACTIVE_OPTIMIZE_WORKERS	Enable CPU-aware worker optimization (Phase 4)	false
REACTIVE_PARALLEL_EXEC	Enable concurrent worker execution	false
CE_INDEX_STATE_STORE	Persist per-file index hashes to .augment-index-state.json	false
CE_SKIP_UNCHANGED_INDEXING	Skip re-indexing unchanged files (requires CE_INDEX_STATE_STORE=true)	false
CE_HASH_NORMALIZE_EOL	Normalize CRLF/LF when hashing (recommended with state store across Windows/Linux)	false
CE_METRICS	Enable in-process metrics collection (Prometheus format)	false
CE_HTTP_METRICS	Expose GET /metrics when running with --http	false
CE_AI_REQUEST_TIMEOUT_MS	Default timeout for AI calls (searchAndAsk) in milliseconds	120000
CE_SEARCH_AND_ASK_QUEUE_MAX	Max queued searchAndAsk requests before rejecting (0 = unlimited)	50
CE_TSC_INCREMENTAL	Enable incremental tsc runs for static analysis	true
CE_TSC_BUILDINFO_DIR	Directory to store tsbuildinfo cache (defaults to OS temp)	(os tmp)
CE_SEMGREP_MAX_FILES	Max files per semgrep invocation before chunking	100
CE_PLAN_AI_REQUEST_TIMEOUT_MS	Timeout for planning AI calls in milliseconds (create_plan, refine_plan, step execution)	300000
CE_HTTP_PLAN_TIMEOUT_MS	HTTP POST /api/v1/plan request timeout in milliseconds	360000

Metrics (optional)

To expose a Prometheus-style endpoint, start the server in HTTP mode and enable both flags:

export CE_METRICS=true
export CE_HTTP_METRICS=true
node dist/index.js --workspace /path/to/project --http --port 3333

Then fetch:

curl http://localhost:3333/metrics

Notes:

• Metrics are intended to use low-cardinality labels (avoid per-query/per-path labels).
• The in-process registry caps total series to prevent unbounded memory growth.

Offline-Only Mode (v1.1.0)

To enforce that no data is sent to remote APIs, set:

export CONTEXT_ENGINE_OFFLINE_ONLY=true

When enabled, the server will fail to start if a remote API URL is configured. This is useful for enterprise environments with strict data locality requirements.

Troubleshooting

Server not showing up in Codex CLI

1. Check ~/.codex/config.toml for syntax errors
2. Ensure paths are absolute
3. Restart Codex CLI
4. Run codex mcp list to see configured servers
5. Use /mcp command in the TUI to check connection status

Authentication errors

Run auggie login or verify environment variables are set correctly.

No search results

Index your workspace first:

node dist/index.js --workspace /path/to/project --index

File watcher not detecting changes (v1.1.0)

1. Ensure you started the server with --watch flag
2. Check that the file is not in .gitignore or .contextignore
3. Wait for the debounce period (default: 500ms) after the last change
4. Check server logs for watcher status messages

Offline-only mode blocking startup (v1.1.0)

If you see an error about offline-only mode:

1. Remove the CONTEXT_ENGINE_OFFLINE_ONLY environment variable, or
2. Configure a localhost API URL in AUGMENT_API_URL

Tool timeout errors during plan generation (v1.4.0)

The create_plan tool can take longer than default MCP client timeouts for complex tasks. If you experience timeout errors, increase the timeout in your MCP client configuration:

For Codex CLI

Edit ~/.codex/config.toml and add or modify the tool_timeout_sec setting under the [mcp_servers.context-engine] section:

[mcp_servers.context-engine]
command = "node"
args = ["/absolute/path/to/context-engine/dist/index.js", "--workspace", "/path/to/your/project"]
tool_timeout_sec = 600  # 10 minutes for complex planning tasks

For Other MCP Clients

Consult your client's documentation for timeout configuration. Common locations:

• Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows)
• Cursor: .cursor/mcp.json in your workspace
• Antigravity: Check client-specific configuration files

Add a timeout setting appropriate for your client's configuration format. A value of 600 seconds (10 minutes) is recommended for complex planning tasks.

Testing

# Run all tests
npm test

# Quieter ESM run (use if you see pipe/stream errors)
node --experimental-vm-modules node_modules/jest/bin/jest.js --runInBand --silent

# Run tests in watch mode
npm run test:watch

# Run tests with coverage
npm run test:coverage

# Interactive MCP testing
npm run inspector

Test Status: 397 tests passing (100% completion) ✅

License

MIT