APX is the reference implementation of the APC protocol. APX is to APC what a language SDK is to a protocol spec.

What APX is

APX is a daemon + CLI that brings the APC convention to life:

• Daemon — a local HTTP server that manages projects, agents, sessions, and message logs
• CLI (apx) — commands for running agents, reading memory, tailing messages, managing sessions
• Web admin — a local web UI served by the daemon to browse projects, agents, sessions, and MCPs from the browser
• Runtimes — bridges to Claude Code, Codex, OpenCode, Aider
• Engines — direct LLM calls via Anthropic, OpenAI, Gemini, Ollama, or a mock
• Plugins — Telegram bot integration out of the box
• MCP support — each agent can expose or consume MCP servers

APX is opinionated about storage: the filesystem is the source of truth. Project definitions and curated memory live in the repo. Runtime state such as sessions, conversations, messages, and caches lives in ~/.apx/ and is never committed.

Quick start

# 1 · Install
npm install -g @agentprojectcontext/apx

# 2 · Set up — interactive wizard (provider → model → channels → daemon)
apx setup

# In any directory with an AGENTS.md, register the project
apx init

# Spawn an agent with a full external runtime
apx run sofia --runtime claude-code "Review the open PRs and summarize them"

# Or a quick one-shot LLM exec
apx exec sofia "What is my role in this project?"

# Watch what's happening
apx messages tail

Examples

Real commands — copy one, point it at an agent like sofia, and APX routes it to the right runtime. The session transcript and its summary land in ~/.apx/; only curated memory and the agent definition live in .apc/.

| What | Command | |------|---------| | Register a project | apx init | | Spawn an agent (full runtime) | apx run sofia --runtime claude-code "Review the open PRs and summarize them" | | Ask a quick question (one-shot) | apx exec sofia "What is my role in this project?" | | Read an agent's memory | apx memory sofia | | Switch runtime, same context | apx run sofia --runtime codex "Add tests for the parser" | | Watch what's happening | apx messages tail |

Use cases

• Review PRs across any runtime — point an agent at your repo; APX routes to Claude Code and falls back to Codex or OpenCode if one isn't installed. The session and its summary land in ~/.apx/.
• Operate your agents from Telegram — talk to project agents from your phone. Identity roles gate who can do what, and every message is logged per channel for a full audit trail.
• Memory that lives in your repo — curated, per-agent memory is plain markdown, committed and reviewable alongside your code. No vendor database, no hidden state, no lock-in.
• Run the same prompt across engines — send one prompt through Anthropic, OpenAI, Gemini or a local Ollama model with apx exec, configured per project or globally.

Installation

npm install -g @agentprojectcontext/apx

Requires Node.js 22+. The daemon starts automatically on first apx call.

Web admin

APX ships a local web admin — the same runtime, in your browser. The daemon serves a single-page app so you can browse and manage everything the CLI does without leaving the UI:

• Projects & agents — see registered projects, open agents, edit roles, models, and skills
• Sessions & messages — read past sessions and tail live activity across every channel
• MCPs, engines & channels — review MCP servers, configure engines, and manage Telegram/desktop

It runs entirely on your machine. Start the daemon (any apx call does this) and open:

apx            # ensures the daemon is up
open http://localhost:7430   # macOS — or just visit it in any browser

The web admin is served from src/interfaces/web/dist at the daemon port (7430 by default, override with APX_PORT). Nothing is sent anywhere — it talks to the local daemon only.

Project layout

Project context — committed to the repository:

project-root/
├── AGENTS.md              ← agent definitions
└── .apc/
    ├── project.json       ← project metadata + stable "id"
    ├── agents/
    │   └── <slug>.md      ← agent definition (role, model, skills…)
    ├── mcps.json          ← MCP servers available to this project
    ├── skills/            ← reusable skill prompts
    └── commands/          ← custom slash commands

Runtime state — local machine only, never committed:

~/.apx/projects/<project-id>/
├── messages/              ← local message history
└── agents/
    ├── <slug>/
    │   ├── sessions/      ← one .md per runtime invocation
    │   └── conversations/ ← LLM conversation threads
    └── default/           ← fallback when no agent role is active
        └── sessions/

Core commands

apx init [path]                          # initialize a project
apx agent list                           # list agents
apx agent add <slug> --role R --model M  # add an agent
apx memory <slug>                        # read agent memory
apx memory <slug> --append "<note>"      # append to memory

apx run   <slug> --runtime claude-code "<prompt>"   # full runtime session
apx run   <slug> --runtime cursor-agent "<prompt>"  # Cursor Agent runtime
apx exec  <slug> "<prompt>"                          # quick LLM call

apx session list <slug>                  # list past sessions
apx messages tail                        # last 50 messages, all channels
apx messages chat --channel telegram     # chat view with user/agent/system type
apx messages tail --channel runtime      # only agent invocations

Message channels

Activity belongs to APX runtime state, not .apc/. Message storage is local to APX, under ~/.apx/:

JSONL messages include type (user, agent, tool, or system) plus actor_id, so chat views can distinguish Telegram users from APX agents and future subagents.

A channel is the surface a turn arrived on. The canonical list lives in src/core/constants/channels.js; voice is a mode, not a channel.

| Channel | What it captures | |---------|-----------------| | cli | apx exec / apx run from the terminal | | telegram | Telegram bot messages | | api | Direct daemon HTTP calls | | web | The admin panel's main chat | | web_sidebar | The panel's side assistant | | web_code | The panel's coding surface | | code | apx code sessions | | deck | The tablet/phone dashboard | | desktop | The floating voice capsule (always voice mode) | | routine | Scheduled routine runs |

Runtimes

| Runtime | Description | |---------|-------------| | claude-code | Spawns Claude Code CLI with the agent's system prompt injected | | codex | OpenAI Codex CLI via non-interactive codex exec --sandbox workspace-write --skip-git-repo-check | | opencode | OpenCode CLI | | aider | Aider CLI | | cursor-agent | Cursor's headless agent | | gemini-cli | Google Gemini CLI | | qwen-code | Qwen Code CLI | | antigravity | Antigravity CLI |

Global APX skill installation also writes named helper skills for codex-cli, claude-code, opencode-cli, and openrouter. They are intentionally narrow and should activate only when those tools/providers are explicitly mentioned.

Engines (for apx exec)

Configured in ~/.apx/config.json:

{
  "engines": {
    "anthropic": { "api_key": "sk-ant-..." },
    "openai":    { "api_key": "sk-..." },
    "ollama":    { "base_url": "http://localhost:11434" },
    "gemini":    { "api_key": "..." }
  }
}

Architecture

APC protocol

APX implements the APC specification. The spec defines the on-disk layout; APX provides the tooling to use it.

License

MIT