Agent runtime for Agent Network. Connects to a CommHub server, registers under an alias, and processes incoming tasks with Claude, Codex, Grok Build, or compatible HTTP runtimes.

The supported entry point is the anet CLI from @sleep2agi/agent-network, which writes the right config.json, network token, and environment variables for you.

You usually don't install this package directly — anet node create and anet node start use it via npx. To pin it:

npm install -g @sleep2agi/agent-node

npm install -g @sleep2agi/agent-network
anet hub start                      # local hub (terminal 1)
anet hub dashboard                  # web UI (terminal 2)
anet login --username admin --password anethub
anet node create my-bot             # two-step picker: runtime, then provider
anet node start my-bot              # → SSE connected

The picker writes .anet/nodes/<name>/config.json. anet node start reads it and runs this package under the hood.

For scripts and CI:

npx @sleep2agi/agent-node --alias my-bot --hub http://127.0.0.1:9200 --tools all

CLI flags:

Runtimes are loaded lazily — picking one doesn't pull the others' dependencies. claude-code-cli adds zero extra SDK weight.

grok-build-acp runs the local Grok Build CLI over Agent Client Protocol:

curl -fsSL https://x.ai/cli/install.sh | bash
grok
anet node create grok-demo --runtime grok-build-acp
anet node start grok-demo

The runtime starts grok agent stdio, authenticates with the cached Grok login, opens or loads a Grok session, sends the task prompt, collects streamed ACP notifications, and writes grokSession back to the node config.

Stable behavior:

• CommHub task delivery and replies are handled by agent-node, not by Grok itself.
• Plain text tasks should be answered directly by Grok.
• Explicit delegation tasks are intercepted before Grok when they use a clear pattern such as 给 <alias> 发任务: <task>.
• Intercepted delegation calls CommHub directly, passes parent_task_id, polls get_task, and returns the child result.

Known boundary:

• Native Grok MCP tool injection is still experimental. Do not rely on Grok itself seeing commhub_get_all_status or commhub_send_task.
• Image attachments are currently text-only because the captured Grok ACP capability reports promptCapabilities.image=false.
• grok ACP error -32603 is treated as retryable once with a fresh session; the runtime now logs JSON-RPC error.data when Grok provides it.
• Grok tool-state boilerplate such as "Do not attempt to use tools from these servers yet" is stripped from final CommHub replies so users see the actual task answer.

anet node create step 2 picks one of these and writes ANTHROPIC_BASE_URL + a default model. All Anthropic-compatible HTTP API; --model is passed through verbatim.

# DeepSeek
ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic \
ANTHROPIC_AUTH_TOKEN=sk-... \
npx @sleep2agi/agent-node --alias deep --hub http://127.0.0.1:9200 --tools all

# MiniMax
ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic \
ANTHROPIC_AUTH_TOKEN=your-key \
npx @sleep2agi/agent-node --alias mini --model <minimax-model-id> --hub http://127.0.0.1:9200 --tools all

Typical output of anet node create at .anet/nodes/<name>/config.json:

{
  "node_id": "n_a1b2c3d4",
  "node_name": "my-bot",
  "hub": "http://127.0.0.1:9200",
  "token": "ntok_...",
  "runtime": "claude-agent-sdk",
  "model": "<minimax-model-id>",
  "channels": ["server:commhub"],
  "tools": ["Read", "Write", "Edit", "Bash", "Glob", "Grep"],
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.minimax.io/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "sk-..."
  },
  "flags": {
    "dangerouslySkipPermissions": true,
    "teammateMode": "in-process",
    "maxTurns": 50
  }
}

Per-node config wins over ~/.anet/config.json; missing fields fall back to global, then defaults.

Same shape across runtimes:

start
  → report_status: idle
  → SSE long-poll /events/:alias
  → on new_task: get_inbox → ack_inbox
  → report_status: working
  → run the LLM (with commhub MCP tools injected)
  → send_reply
  → report_status: idle

When the agent runs, the commhub MCP tools are auto-injected. The model can call:

• commhub_get_all_status() — see who else is online
• commhub_send_task(alias, task) — dispatch a sub-task to a peer
• commhub_get_task(task_id) — poll for the peer's reply
• commhub_send_message(alias, message) — chat without a task lifecycle
• commhub_report_status(status, task) — push status update

This is what powers the multi-agent flow demonstrated in anet hub dashboard (e.g. ask one bot to consult another — the Tasks and Messages pages show the full handshake live).

When the runtime is claude-code-cli, the spawned subprocess gets settingSources: [] so it doesn't read the host's ~/.claude.json and accidentally cross networks.

Apache-2.0