amcl-server 1.3.3

A/MCL — Agent/Multi-Coding-agent Context Layer. MCP server for automatic context persistence across AI coding agents.

A/MCL — Agent/Multi-Coding-agent Context Layer

Zero-intervention context persistence across AI coding agents.

When you hit a rate limit on one agent (Codex, Antigravity, Cursor, Claude Code) and switch to another, A/MCL ensures the new agent automatically has access to the complete conversation history, file changes, reasoning chains, and project state. No commands. No manual handoff. It just works.

---

How It Works

A/MCL is an MCP (Model Context Protocol) server. MCP-compatible agents discover it automatically at launch and gain access to shared project context via standard tools, resources, and prompts.

┌──────────────┐     stdio      ┌────────────────┐     SQLite     ┌──────────┐
│  AI Agent    │ ◄────────────► │  A/MCL Server  │ ◄────────────► │  Context │
│ (Codex,      │                │  (FastMCP)     │                │  Store   │
│  Cursor, etc)│                │                │                │ (~/.amcl)│
└──────────────┘                └────────────────┘                └──────────┘

1. Install once → pip install amcl-server (or pip3 install)
2. Setup → amcl-server setup (creates DB, shows MCP config)
3. Use your agents normally → Context is automatically shared
4. Switch agents → New agent picks up exactly where the last one left off

---

Quick Start

# Install universally
pip install amcl-server  # or pip3 install amcl-server

# Initialize and Auto-Register
amcl-server setup

# Verify integrations
amcl-server check

# Check status
amcl-server status

Upgrading an Existing Installation

pip install --upgrade amcl-server
amcl-server setup
amcl-server check

The setup command automatically detects all installed AI agents and IDE extensions on your system and registers A/MCL directly into their settings. Once you run setup, you are completely done. It also installs or updates Codex's global ~/.codex/AGENTS.md so Codex inherits the same A/MCL behavior rules without touching repository-local AGENTS.md files.

Optional: more accurate token counting

A/MCL ships with a zero-dependency token estimator that matches GPT-4's tiktoken cl100k_base within ~7% mean absolute error (max ~13%) for code, prose, JSON, markdown, identifiers, and unicode. It's a pure stdlib re-based heuristic — no model files, no install, no deps.

For exact token counts:

pip install amcl-server[accurate-tokens]

This adds tiktoken (~2.7 MB) for cl100k_base-precise tokenization. A/MCL uses it automatically when available and transparently falls back to the built-in heuristic otherwise.

---

Supported Agents (Auto-Detected)

A/MCL automatically integrates with:

• Cursor (~/.cursor/mcp.json)
• Codex (~/.codex/config.toml)
• Claude Code (~/.claude.json user scope + project .mcp.json)
• Claude Desktop (Library/Application Support/Claude/claude_desktop_config.json)
• Antigravity (~/.gemini/antigravity/mcp_config.json)
• Amp (~/.amp/mcp.json)
• Roo / Cline (VSCode & Cursor instances)
• Generic MCP clients (~/.mcp/config.json)

If your agent isn't automatically found, the setup command will print the exact JSON snippet you can manually paste into its MCP configuration.

---

What Gets Shared

Category	Data
Conversation	All messages between user and agents, with agent attribution
File Changes	Files created, modified, deleted — tracked automatically
Tasks	Current goal, active tasks, completed work, blockers
Decisions	Design decisions with rationale and alternatives considered
Agent History	Which agents worked on what, session timestamps

---

MCP Tools

Agents can invoke these tools:

Tool	Description
context_get_current	Full project context snapshot
context_update	Append new conversation/file/decision data
context_query	Search context history
context_get_files_changed	Files changed since last agent switch
context_get_conversation	Recent N messages
context_get_reasoning	Decision log
context_add_decision	Log a design decision
context_mark_complete	Mark a task as completed
context_add_blocker	Record a blocking issue

---

MCP Resources

Read-only data endpoints:

• context://current-project — Complete snapshot
• context://conversation — Message history
• context://files — File state & changes
• context://reasoning — Decision log
• context://tasks — Task state
• context://agents — Agent session history

---

MCP Prompts

Pre-built prompts for context injection:

• Project Context — Complete project overview
• Recent Work — Summary of last session
• Continuation — "Pick up where we left off"
• Decisions Log — Key decisions and rationale

---

Environment Variables

Variable	Default	Description
AMCL_DATA_DIR	~/.amcl	Where context database lives
AMCL_LOG_LEVEL	WARNING	Logging verbosity
AMCL_AGENT_NAME	unknown	Name of the connecting agent
AMCL_PROJECT_DIR	cwd	Project directory override

---

---

Privacy

All data stays local on your machine in ~/.amcl/amcl.db. Nothing is ever sent to external servers. No telemetry, no analytics.

---

Contributing

Bug reports, docs, and pull requests are welcome. See CONTRIBUTING.md for development setup and guidelines.

Please read SECURITY.md before reporting security issues publicly. See CHANGELOG.md for release notes.

License

This project is licensed under the MIT License. See LICENSE for the full text.