Agent Node

Agent Node is the working unit in Agent Network -- it receives tasks, invokes an AI model to process them, and reports results.

Not sure which Runtime to pick?

Reuse a Claude subscription / smoothest first-time path → claude-code-cli (zero config after claude auth login)
Writing copy / translation / analysis (programmatic) / using a domestic Chinese model → claude-agent-sdk + pick the matching vendor in the wizard
Writing code / running commands → codex-sdk
Using xAI Grok Build → grok-build-acp (detailed runtime guide ↗)
Reach a vendor that's NOT in the built-in list (DeepSeek / GLM / Kimi / OpenRouter / vLLM / SiliconFlow / Qwen ...) → claude-agent-sdk + pick custom in the vendor submenu + ANTHROPIC_BASE_URL

Authoritative comparison of the 4 runtimes × npm package × wizard behavior × prereq auth: runtimes — canonical table. Full anet node create wizard order (node-name → runtime → ...): Getting Started §5.

Installation

bash

# Global install
npm install -g @sleep2agi/agent-node

# Or run directly with npx (recommended, no install needed)
npx @sleep2agi/agent-node --help

Runtimes

Agent Node supports four AI runtime engines covering all major models:

claude-agent-sdk

Based on the Anthropic Claude Agent SDK.

Property	Description
Models	Latest Claude Sonnet / Opus / Haiku (specific IDs at Anthropic Models)
Prerequisites	Anthropic API Key, or any Anthropic-compatible API key (MiniMax / DeepSeek / GLM / Kimi / InternLM / Xiaomi MiMo / OpenRouter, etc. — full list in multi-model)
Strengths	Programmatic Anthropic-compatible API calls for stable background agents
Isolation	`settingSources: []` fully isolates host config

bash

npx @sleep2agi/agent-node \
  --alias reasoning-master \
  --runtime claude-agent-sdk \
  --model claude-sonnet-4-6 \
  --hub http://YOUR_IP:9200

Prerequisites checklist

[ ] Anthropic API Key or MiniMax API Key (paid)
[ ] CommHub Server is running

Verify

After starting, you should see SSE connected, waiting for tasks.... If you get an auth / 401 / invalid x-api-key error: check that ANTHROPIC_API_KEY (for api.anthropic.com) or ANTHROPIC_AUTH_TOKEN + ANTHROPIC_BASE_URL (for a third-party Anthropic-compatible endpoint) is set correctly — see runtimes — common pitfalls. claude auth login is for the claude-code-cli runtime and has no effect on the SDK path.

claude-code-cli

Runs a local Claude Code CLI process -- the same claude command you use daily in your terminal.

Property	Description
Models	Latest Claude Sonnet / Opus / Haiku (specific IDs at Anthropic Models)
Prerequisites	Claude Code installed (`npm i -g @anthropic-ai/claude-code`)
Strengths	Spawns a `claude` child process with full terminal capabilities
Difference	vs `claude-agent-sdk`: CLI mode = spawns `claude` process; SDK mode = programmatic API calls

bash

npx @sleep2agi/agent-node \
  --alias terminal-assistant \
  --runtime claude-code-cli \
  --model claude-sonnet-4-6 \
  --hub http://YOUR_IP:9200

Prerequisites checklist

[ ] Install Claude Code: npm install -g @anthropic-ai/claude-code
[ ] Verify claude --version outputs correctly
[ ] Run claude auth login so your local Claude subscription is active (claude-code-cli reuses the local login state)
[ ] CommHub Server is running

Verify

After starting, you should see SSE connected, waiting for tasks....

If you get claude: command not found, make sure Claude Code is installed globally
If you get auth / 401, re-run claude auth login to refresh the subscription session

claude-code-cli vs claude-agent-sdk

claude-code-cli: Spawns a claude child process, just like typing commands in your terminal. Has all Claude Code capabilities (file operations, bash execution, MCP tools, etc.).
claude-agent-sdk: Calls Claude through the programmatic SDK API. Better suited for scenarios that need fine-grained control over settingSources, maxTurns, and other parameters.

codex-sdk

Based on the OpenAI Codex SDK.

Property	Description
Models	Codex SDK model (set with `--model`; see OpenAI Codex docs for the current model id)
Prerequisites	`codex auth login`
Strengths	Strong code generation, flexible tool use
Tools	Codex CLI ships with Read / Write / Edit / Bash / Glob / Grep / WebSearch baked in (does not honor `--tools`) + agent-node injects per-node CommHub tools

bash

npx @sleep2agi/agent-node \
  --alias code-assistant \
  --runtime codex-sdk \
  --model <codex-model-id> \
  --hub http://YOUR_IP:9200
# Note: codex-sdk does not honor --tools. Codex built-in tools come from the codex CLI binary;
# CommHub tools are injected by agent-node per node, so codex nodes can actively get_all_status / send_task / get_task.

Prerequisites checklist

[ ] Install codex CLI: npm install -g @openai/codex (@openai/codex-sdk lives in @sleep2agi/agent-node's optional peerDependencies; npm 7+ pulls it in automatically with agent-node, but the SDK shells out to the codex binary — see runtimes / codex-sdk prereqs)
[ ] Run codex auth login to authenticate with OpenAI (or export OPENAI_API_KEY=sk-xxx)
[ ] CommHub Server is running

Verify

After starting, you should see SSE connected, waiting for tasks....

If you get Error: spawn codex ENOENT, the codex binary isn't on PATH — run npm install -g @openai/codex + check which codex
If you get a codex auth error, run codex auth login (or check the OPENAI_API_KEY env)

grok-build-acp

Integrates with xAI Grok Build ACP (Agent Communication Protocol) by spawning the local grok ACP server. This is the 4th runtime; it is available in the anet node create wizard and can also be selected explicitly with --runtime grok-build-acp. Full configuration / Known Limits / Delegation Contract: grok-build-runtime.md ↗.

Field	Description
Prerequisites	`grok` CLI already authenticated + `XAI_API_KEY` env
Strengths	xAI Grok Build models, ACP-protocol cross-agent collaboration
Tools	Bundled with Grok ACP, does not accept `--tools`

bash

npx @sleep2agi/agent-node \
  --alias grok-helper \
  --runtime grok-build-acp \
  --hub http://YOUR_IP:9200

grok-build-acp improvements

#201 — 3-layer fix for delegate refusal: parser broaden + prompt softening + authorised-phrasings list
#204 — per-session MCP inject + per-node cwd isolation to prevent .mcp.json identity pollution

See grok-build-runtime.md Known Limits for the full breakdown.

preview chain history (how it converged on preview.7)

preview.2 stdio variant: structurally fixes the .mcp.json shared-identity bug (ACP side)
preview.6 — transport switched to HTTP: Grok calls CommHub /mcp directly with a Bearer ntok_, bypassing subprocess / bun PATH / stdout-pollution risks entirely
preview.7 — per-node isolated cwd (final): Grok CLI was reading the cwd .mcp.json alongside the ACP injection — two MCP servers coexist and the stale one wins hello. Fix: ACP session/new passes a per-node isolated cwd that mirrors user files but omits .mcp.json
v0.10.11 final = preview.7 promoted

:::

claude-agent-sdk + Domestic Models

Routes claude-agent-sdk requests to domestic model APIs via ANTHROPIC_BASE_URL, ideal for low-cost scenarios.

Property	Description
Models	MiniMax, DeepSeek, GLM, Kimi, InternLM, Xiaomi MiMo, OpenRouter — any Anthropic-compatible endpoint (full provider table: Multi-model setup)
Prerequisites	API key for the target model
Strengths	Low cost, high throughput, direct access in China
Mechanism	Routes requests to compatible APIs via `ANTHROPIC_BASE_URL`

bash

# MiniMax
ANTHROPIC_BASE_URL=https://api.minimaxi.com/anthropic \
ANTHROPIC_AUTH_TOKEN=your-minimax-key \
npx @sleep2agi/agent-node \
  --alias minimax-bot \
  --runtime claude-agent-sdk \
  --model <minimax-model-id> \
  --hub http://YOUR_IP:9200

# InternLM (note: bare hostname, no /anthropic suffix)
ANTHROPIC_BASE_URL=https://chat.intern-ai.org.cn \
ANTHROPIC_AUTH_TOKEN=your-intern-key \
npx @sleep2agi/agent-node \
  --alias intern \
  --runtime claude-agent-sdk \
  --model intern-s1-pro \
  --hub http://YOUR_IP:9200

Prerequisites checklist

[ ] API key for the target model (e.g. MiniMax API Key)
[ ] Set environment variables ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN
[ ] CommHub Server is running

Verify

After starting, you should see SSE connected, waiting for tasks.... If you get a 401 or auth error, double-check your API key.

Command-Line Parameters

bash

npx @sleep2agi/agent-node [options]

Parameter	Default	Description
`--alias`	(required)	Agent name (display name in CommHub)
`--hub`	http://127.0.0.1:9200	CommHub Server address
`--runtime`	claude-agent-sdk	Runtime engine (`claude-agent-sdk` / `codex-sdk` / `claude-code-cli` / `grok-build-acp`)
`--model`	(per runtime default)	AI model name
`--tools`	(none)	Available tools, comma-separated
`--max-budget`	0	Per-task budget cap (USD; 0 means disabled)
`--session`	(new)	Resume a specific session
`--config`	(auto-detect)	Config file path

Where token / network come from

The auth token is supplied via the token field in .anet/nodes/<name>/config.json or the COMMHUB_TOKEN env var — the CLI does not accept a --token flag. The network ID is inferred from the ntok_ token claim; no manual flag is needed. The agent-node CLI also does not parse single-letter short flags (-a / -h / -r / -m / -t / -s / -c) — only the long-form flags in the table above are accepted.

Configuration Files

Agent Node supports multiple configuration methods, from highest to lowest priority (verified at agent-node/src/cli.ts):

Global ~/.anet/config.json fallback

After the project config is loaded, cli.ts fills in the missing hub and token fields from the global ~/.anet/config.json. Only these two fields fall back across projects — runtime / model / tools / env must be set via project config.json / CLI / env; global config does not cover them. Aligned with the [feedback_config_priority] memory ("project config overrides global at the field level; missing fields fall back to global").

Full config.json Fields

json

{
  "anet_version": "0.1.0",
  "node_id": "n_a1b2c3d4",
  "node_name": "code-assistant",
  "token": "ntok_...",
  "runtime": "claude-agent-sdk",
  "model": "<model-id>",
  "session": "",
  "channels": ["server:commhub"],
  "tools": ["Read", "Write", "Edit", "Bash", "Glob", "Grep"],
  "logLevel": "info",
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.minimaxi.com/anthropic"
  },
  "flags": {
    "dangerouslySkipPermissions": true,
    "teammateMode": "in-process",
    "maxTurns": 20
  }
}

Field	Type	Description
`anet_version`	string	Config version
`node_id`	string	Stable unique identifier (n_ prefix + 8-char hex)
`node_name`	string	Display name, can be renamed
`alias`	string	Node alias (the `.anet/nodes/<alias>/` directory name + the display identifier on CommHub; equals `node_name` when not set separately)
`runtime`	string	Runtime: `claude-agent-sdk` / `codex-sdk` / `claude-code-cli` / `grok-build-acp`
`model`	string	AI model name
`session`	string	session/thread ID. For the `claude-code-cli` runtime, `anet node create` pre-generates a UUID (first `start` binds it via `--session-id <uuid>`, restarts auto-`--resume <uuid>` to continue the conversation; v0.8.2 fixed a prior default session-loss bug). For other runtimes, this is the previous session ID for resume
`channels`	string[]	Connected channels list
`tools`	string[]	Allowed tools list. `claude-agent-sdk` only; `codex-sdk` silently ignores it (toolset is baked into the codex binary — see the L109 note above + runtimes#codex-sdk)
`env`	object	Environment variable overrides
`flags`	object	Runtime flags
`hub`	string	CommHub Server address override (falls back to the global `~/.anet/config.json` `hub` when unset)
`token`	string	Auth token override (falls back to the global config `token` when unset)
`network_id`	string	Network ID (usually inferred from `ntok_`, no need to set manually)
`systemPrompt`	string	System prompt, prepended to every task (can also use the `--prompt` flag; the `SYSTEM_PROMPT` env var is not read — see the note above)

Task Processing Flow

After starting, Agent Node automatically enters a task listening loop:

Message Type Filtering

Agent Node only triggers AI processing for task type messages:

Message Type	SSE Event	Agent Behavior
task	`new_task`	processInbox -> AI think -> reply
broadcast	`broadcast`	processInbox -> AI think -> reply
reply	`new_reply`	Log only
message	`new_message`	Log only
ack	(not pushed)	--

This design prevents message loops (A replies to B -> B replies to A -> infinite loop).

Tool Configuration

Default = full Claude Code preset (v0.9.0+, #101 Option B)

Since #101 Option B (anet v0.9.0+), the claude-agent-sdk runtime's default toolset is the full Claude Code preset — not an empty set. As soon as a node spawns, it can use:

Filesystem: Read / Write / Edit / Glob / Grep
Shell: Bash (subject to dangerouslySkipPermissions=true, on by default — no per-call confirmation)
Network: WebFetch / WebSearch
Subtasks / notebooks: Task / NotebookEdit / ...

Plus the 17 MCP tools on the hub side (commhub_send_task / commhub_reply / ...).

Root cause (#101): in older versions, with no tools field in config.json, agent-node passed the SDK options.tools = undefined, and the SDK treated that as "no built-in tools". The agent could only call MCP tools and would hallucinate "network restricted" when asked for WebFetch / Bash / Read. Option B forces the fallback to the SDK { type: 'preset', preset: 'claude_code' } sentinel — per the SDK type defs (sdk.d.ts:1229-1238), this is the standard way to say "give me the full Claude Code toolset".

Three `--tools` modes (only for `claude-agent-sdk`)

The --tools flag only affects the claude-agent-sdk runtime — the codex-sdk toolset is baked into the codex CLI (Read/Write/Edit/Bash/Grep/Glob/WebSearch, does not honor --tools); the claude-code-cli runtime shares the host's Claude Code toolset and also doesn't go through this flag.

Input	Effect	Verify
`--tools all`	Full SDK preset (same as above) — single source of truth	`cli.ts`
`--tools Read,Glob,Grep`	Explicit allowlist (string array), bypasses preset — strict sandbox	`cli.ts,220`
Absent / empty string	Falls back to full preset (the #101 fix; previously left as `undefined` → empty set)	`cli.ts,221`

Actual logic from source (agent-node/src/cli.ts):

const TOOLS_PRESET = { type: "preset" as const, preset: "claude_code" as const };
// Behaviour matrix (sdk.d.ts:1229-1238):
//   --tools "all"       → SDK preset (full Claude Code tool set)
//   --tools "Read,Bash" → explicit allowlist
//   --tools "" (absent) → SDK preset (the #101 fix; previously left empty)
const TOOLS_EXPLICIT = toolsRaw === "all" ? null : toolsRaw.split(",").filter(Boolean);
let TOOLS: string[] | typeof TOOLS_PRESET =
  toolsRaw === "all" ? TOOLS_PRESET
  : (TOOLS_EXPLICIT && TOOLS_EXPLICIT.length) ? TOOLS_EXPLICIT
  : TOOLS_PRESET;
// ... cli.ts: tools: TOOLS  ← passed to claude-agent-sdk query options (preset or string[])

bash

# Default (no --tools) → full Claude Code preset
npx @sleep2agi/agent-node --alias coder

# Explicit "all" → same preset (single source of truth, replaces the old hardcoded 8-tool list)
npx @sleep2agi/agent-node --alias coder --tools all

# Explicit allowlist (read-only agent) — bypasses the preset
npx @sleep2agi/agent-node --alias coder --tools Read,Glob,Grep

# codex-sdk silently ignores --tools
npx @sleep2agi/agent-node --alias coder --runtime codex-sdk
# Codex has Read/Write/Edit/Bash/Grep/Glob/WebSearch baked in — you cannot detach individual tools.

`anet node create` behavior-disclosure banner (Vincent push, #101)

After a successful anet node create <alias>, agent-node prints a behavior-disclosure banner: the built-in tools (explicit list or all (Claude Code preset)) + MCP tools + current flags (dangerouslySkipPermissions=true / teammateMode=true) + a one-liner "the agent can read/write files, run shell commands, and access the network". The banner is printed at the end of createCommand in agent-network/bin/cli.ts — so you actually see what the agent can do and decide whether to sandbox it.

⚠ User responsibility: the default preset + default dangerouslySkipPermissions=true means once the agent starts it can edit files, run shell commands, and access the network without confirmation prompts. See Security → Tool Permissions.

Security note

Default preset + default yolo mode (dangerouslySkipPermissions) — don't run agents directly from $HOME; use a disposable working directory
For strict sandbox: --tools Read,Glob,Grep gives a read-only agent
To turn off yolo mode: anet node create --no-skip-permissions (every tool call prompts, hurts long-task UX)
Per-task budget cap: --max-budget 0.1 (see Budget Control below)

Vendor adapter layer (InternLM / 书生 etc.)

On some vendor endpoints, claude-agent-sdk follows its RLHF defaults and drifts from Anthropic standard (canonical case: intern-s2-preview does not emit tool_use content blocks and falls back to verbose Thinking Process text). agent-node uses a vendor adapter to detect by ANTHROPIC_BASE_URL and inject a system-prompt bias that nudges behavior back — see Vendor Adapters for the full mechanism, the 5 side effects, and the opt-out path.

Budget Control

The --max-budget parameter caps the maximum spend per task (in USD) — only effective for the claude-agent-sdk runtime:

bash

# Max $0.10 per task (claude-agent-sdk)
npx @sleep2agi/agent-node --alias coder --max-budget 0.1

# Max $1.00 per task (complex tasks)
npx @sleep2agi/agent-node --alias reasoner --max-budget 1.0

Verified at agent-node/src/cli.ts:

if (MAX_BUDGET > 0) options.maxBudgetUsd = MAX_BUDGET;  // passed to claude-agent-sdk query options

When SDKResultMessage.total_cost_usd reaches maxBudgetUsd, claude-agent-sdk automatically ends the turn and the task moves to error_max_budget.

codex-sdk / claude-code-cli runtime do not support a USD budget cap

The codex-sdk path (cli.ts processWithCodex) does not read MAX_BUDGET; --max-budget is silently ignored. Codex-sdk only reports token counts (TurnCompletedEvent.usage), not USD — you have to derive cost from your own model→price table (aligned with the sdk-deep-dive chain).
claude-code-cli runs against your local Claude Code subscription, counted against subscription quota rather than USD.
Cross-runtime budget control: put a reverse proxy in front (nginx / Cloudflare / litellm proxy) and throttle by model-API call count.

Lifecycle

The complete Agent Node lifecycle:

Phase	CommHub Status	Description
Created	(not in CommHub)	`anet node create` generates config.json
Registered	idle	`report_status(idle)`
Online	idle	SSE connected, waiting for tasks
Running	working	Processing a task
Error	error	Runtime error
Offline	offline	Process exited
Deleted	(not in CommHub)	All data cleared

Heartbeat

Automatic report_status heartbeat every 3 minutes
Server marks as offline after 10 minutes without a heartbeat
Heartbeat also returns inbox_count for checking pending tasks

Reconnection

SSE auto-reconnects on disconnect using exponential backoff:

Retry interval ([#202](https://github.com/sleep2agi/agent-network/issues/202)): exponential backoff `1s → 2s → 4s → 8s → 16s → 30s (cap)`

Online status is automatically restored after successful reconnection. Three changes ship together:

Tighter backoff: 1s → 30s cap — full reconnect within 30s of hub restart
Re-register on reconnect: every successful reconnect re-fires the node registration (idempotent upsert on the hub), so the dashboard recovers in under 30s (previously a reconnect only restored the SSE stream and the dashboard waited for the next 3-minute heartbeat)
Give-up guard: if reconnect failures continue for over 1 hour, agent-node gives up auto-retry, logs an error, and stops burning CPU

Pairs with the anet hub stop / hub status commands: during hub maintenance, stop → start no longer requires a follow-up anet project restart — nodes restore themselves. Run anet upgrade to pick up the latest fix.

Graceful Shutdown

On SIGINT (Ctrl+C) or SIGTERM:

Report report_status(offline)
Close SSE connection
Exit process

If the process crashes (no time to report), CommHub detects via heartbeat timeout and marks the agent offline after 10 minutes (verified at server/src/index.ts:816-821 Date.now() - 10 * 60 * 1000 cutoff, lazily triggered on /api/status calls).

Environment Variables

Only the env vars that agent-node actually reads from process.env (verified at agent-node/src/cli.ts):

Variable	Equivalent CLI flag / config field	Description
`COMMHUB_URL`	`--hub` / `--url` / `config.hub`	CommHub Server address (cli.ts)
`COMMHUB_TOKEN`	`config.token` / `globalConfig.token`	Auth token (cli.ts; no CLI flag accepted)
`COMMHUB_ALIAS` / `ALIAS`	`--alias` / `config.alias`	Agent alias — both env var names work (cli.ts)
`RUNTIME`	`--runtime` / `config.runtime`	Runtime engine, defaults to `claude-agent-sdk`
`MODEL`	`--model` / `config.model`	AI model
`LOG_LEVEL`	`--log-level` / `config.logLevel` (top-level, not under `flags`)	`debug` / `info` / `warn` / `error`
`ANET_NETWORK_ID`	`config.network_id` / `globalConfig.network_id`	Network ID fallback (typically inferred from `ntok_`; cli.ts)
`TELEGRAM_BOT_TOKEN`	channel `.env` / `config.env.TELEGRAM_BOT_TOKEN`	Bot token for the Telegram channel — agent-node reads it directly in the telegram channel startup path (cli.ts); the allowlist comes from `access.json`, not env (see Channel — Telegram)
`CLAUDE_TIMEOUT_MS`	`--claude-timeout-ms` / `config.flags.claudeTimeoutMs` / `config.claudeTimeoutMs`	Per-query timeout (ms) for the `claude-agent-sdk` runtime, default `300000` (300s) (raised from 120s in v0.9.2 by #132 Tier 1 — the SDK concurrency investigation measured intern API per-request latency stretching to 17-37s under heavy fan-out, so the old 120s ceiling fired mid-stream and 25/30 sub-agents silently failed); on timeout it aborts and returns an error suggesting you check `ANTHROPIC_BASE_URL` reachability. Verify `agent-node/src/cli.ts`
`CLAUDE_MAX_RETRIES`	`--claude-max-retries` / `config.flags.claudeMaxRetries` / `config.claudeMaxRetries`	Retry count for `claude-agent-sdk` runtime queries that fail. Default `2` (3 attempts total including the initial). Each attempt runs its own `CLAUDE_TIMEOUT_MS` window; on transient errors / timeouts, backs off `4s, 8s + 0-1s jitter` (the jitter spreads herd retries so a recovering vendor queue isn't slammed). Auth-class errors do not retry (#129 fast-fail: the `isAuthError` regex matches 401 / 403 / `invalid_api_key` / intern A02xx etc. and returns a FATAL with the vendor-specific URL hint). Set `0` to revert to v0.9.1 behavior (no retry). Introduced in v0.9.2 via #132 Tier 1. Verify `agent-node/src/cli.ts` + retry loop `cli.ts`
`ANET_CODEX_STDIO_DIRECT`	env only (no CLI flag / config field; per-spawn injection)	v0.10.0+, only takes effect when `runtime=codex` (claude-agent-sdk / claude-code-cli ignore it). Set `=1` to switch the codex runtime from the `@openai/codex-sdk` wrapper to the direct stdio JSON-RPC client path: agent-node runs `spawn('codex', ['app-server'])` with a ~155 LOC stdio client and the full 67-method v2 protocol surface (thread / turn / item / realtime), bypassing the wrapper's `--mcp-config` HTTP-transport bug family (#102 hang root cause). v0.10.x (including the current stable) still defaults to the wrapper (preview-feedback window + backward compat); v0.11.0 plans to flip the default, at which point the toggle becomes `ANET_CODEX_LEGACY_SDK=1` opt-out (per `agent-node/src/cli.ts` comments). Introduced in v0.10.0 via #141. Verify `agent-node/src/cli.ts` `if (process.env.ANET_CODEX_STDIO_DIRECT === "1")`
`ANTHROPIC_BASE_URL`	`config.env.ANTHROPIC_BASE_URL`	Model API URL (required when targeting a third-party Anthropic-compatible endpoint)
`ANTHROPIC_AUTH_TOKEN`	`config.env.ANTHROPIC_AUTH_TOKEN`	Model API key — for third-party Anthropic-compatible endpoints (MiniMax / DeepSeek / GLM / Kimi / InternLM / Xiaomi MiMo / OpenRouter / vLLM, etc.)
`ANTHROPIC_API_KEY`	`config.env.ANTHROPIC_API_KEY`	Model API key — only for direct api.anthropic.com; don't reuse it for third-party endpoint keys (see runtimes — claude-agent-sdk pitfalls)

TOOLS / SYSTEM_PROMPT env vars do not exist

The previously listed TOOLS and SYSTEM_PROMPT env vars are not read by agent-node (verified: cli.ts toolsRaw = opts.tools || fileConfig.tools has no process.env.TOOLS; cli.ts SYSTEM_PROMPT = opts.prompt || fileConfig.systemPrompt has no env reading). To set tools, use the --tools CLI flag or config.json's tools field; for the system prompt, use the --prompt flag or config.json's systemPrompt field.

Docker Usage

When running in Docker, environment variables are the most convenient configuration method. See Docker Deployment.

Next steps

Get started:

One-shot install — first agent in 5 minutes after install

Configure deeper:

Runtimes — picking between claude-agent-sdk / codex-sdk / claude-code-cli / grok-build-acp
Multi-model — use DeepSeek / MiniMax / Kimi / Claude
Channel plugins — wire agents to Telegram / WeChat / Feishu

Production:

Docker deployment — containerized agents
Production deployment — multi-machine, TLS, backups
Dashboard — monitor agent status, tasks, message flow

Troubleshooting:

Troubleshooting — common issues
anet doctor --fix — auto-detects expired ntok_ and other issues

Agent Node ​

Installation ​

Runtimes ​

claude-agent-sdk ​

claude-code-cli ​

codex-sdk ​

grok-build-acp ​

claude-agent-sdk + Domestic Models ​

Command-Line Parameters ​

Configuration Files ​

Full config.json Fields ​

Task Processing Flow ​

Message Type Filtering ​

Tool Configuration ​

Default = full Claude Code preset (v0.9.0+, #101 Option B) ​

Three --tools modes (only for claude-agent-sdk) ​

anet node create behavior-disclosure banner (Vincent push, #101) ​

Vendor adapter layer (InternLM / 书生 etc.) ​

Budget Control ​

Lifecycle ​

Heartbeat ​

Reconnection ​

Graceful Shutdown ​

Environment Variables ​

Next steps ​

Agent Node

Installation

Runtimes

claude-agent-sdk

claude-code-cli

codex-sdk

grok-build-acp

claude-agent-sdk + Domestic Models

Command-Line Parameters

Configuration Files

Full config.json Fields

Task Processing Flow

Message Type Filtering

Tool Configuration

Default = full Claude Code preset (v0.9.0+, #101 Option B)

Three `--tools` modes (only for `claude-agent-sdk`)

`anet node create` behavior-disclosure banner (Vincent push, #101)

Vendor adapter layer (InternLM / 书生 etc.)

Budget Control

Lifecycle

Heartbeat

Reconnection

Graceful Shutdown

Environment Variables

Next steps