Agent Context Server

A portable, tool-agnostic memory and code structure graph for agentic AI coding tools (OpenCode, Antigravity, Claude Code, etc.).

🧠 The Problem

Agentic AI tools typically keep their memory locked inside a single session. When you run out of context tokens or switch to a different AI tool, you lose:

• What you were currently working on (tasks, pending steps).
• Why past technical decisions were made.
• The structural architecture of the codebase.

🚀 The Solution

The Agent Context Server is an MCP (Model Context Protocol) server that tracks working context and code structure in a lightweight, local SQLite database (.agentctx/context.db). It provides a single, portable memory that follows the project rather than the tool.

You can start a task in one AI tool, hit a context limit, open a completely different AI tool, and resume exactly where you left off.

✨ Features

• Working Context Engine: Tracks active tasks, decisions, progress logs, and open questions across sessions.
• Code Structure Graph: Uses tree-sitter to parse and map Python, JavaScript, and TypeScript files, extracting classes, functions, and methods without sending your code to external APIs.
• Incremental Indexing: Smart content-hashing ensures only modified files are re-indexed.
• Project Export: Built-in tools to instantly zip and backup your agent's memory state.
• Fully Local & Private: No LLM calls inside the server; all state is kept entirely locally via SQLite.

🛠️ Prerequisites

• Python 3.10+
• uv (Fast Python package and project manager)

📦 Installation

1. Clone this repository to your machine:

   git clone (replace with your own repo URL once you push this to GitHub)
   cd agent-context-mcp-server
   

2. The project uses uv for dependency management. The required packages (mcp, tree-sitter, tree-sitter-python, tree-sitter-javascript, tree-sitter-typescript) will be automatically managed when you run the server.

⚙️ MCP Configuration

To use this server with your preferred agentic AI tool (like Google Antigravity, OpenCode, or Claude Code), add the following to your MCP configuration settings:

{
  "mcpServers": {
    "agent-context-server": {
      "command": "uv",
      "args": [
        "--directory",
        "/absolute/path/to/agent-context-mcp-server",
        "run",
        "server.py"
      ]
    }
  }
}

Note: Replace /absolute/path/to/agent-context-mcp-server with the actual path where you cloned this repository. The server will dynamically create the .agentctx/ directory in whatever project folder your AI agent is currently working in.

---

🧰 Available Tools

Once connected, your AI agent will have access to the following tools:

Working Context

• start_task(title): Creates a new working context task.
• log_progress(task_id, entry, entry_type): Appends a progress log (step_done, error_hit, next_step, note).
• log_decision(task_id, decision, reason, symbol_ref): Logs technical decisions and the reasoning behind them.
• add_question(question) / resolve_question(question_id): Manages open questions.
• resume(task_id?): Retrieves the full state of the active task to seamlessly pick up work.
• list_tasks(status?): Lists active, done, blocked, or abandoned tasks.

Structure Graph

• graph.index(path, languages?): Parses a file (.py, .js, .ts, .jsx, .tsx) and extracts structural symbols.
• graph.get_symbol(name): Retrieves a specific symbol's file path and line numbers.
• graph.trace_calls(symbol_name, direction): Traces inbound or outbound function calls within the same file.
• graph.get_architecture(): Returns a high-level overview of the parsed codebase, including files, symbol counts, and root entry points.

Lifecycle

• project.export(): Zips the entire .agentctx/ directory for safe backup and sharing.

---

📄 License

• MIT License