ZeroClaw Setup: MiniMax M3, GLM-5.3 & Discord on Your VPS
Updated ZeroClaw setup guide for v0.8: install the self-hosted AI assistant on a VPS, configure MiniMax M3 and GLM-5.3, add Discord, and secure your bot.
Updated Published 26 min read

Disclosure: this guide contains referral links. If you buy through them, I may earn a commission at no extra cost to you.
A proper ZeroClaw setup on a VPS gives you a self-hosted AI assistant written in Rust that boots in milliseconds and uses less memory than a browser tab. ZeroClaw is now at v0.8.7 (October 2026) and the project has moved fast since I first covered it in February - the install flow, config schema, and recommended models have all changed. This guide walks through the current setup with MiniMax M3 and GLM-5.3 as your models and Discord as the chat channel.
Updated October 2026 - rewritten for ZeroClaw v0.8.x. If you followed the February 2026 version of this guide, the onboard command is gone, the config format changed to Schema V3, and the models are two generations newer. The “What’s new” section below covers what broke.
If you’re comparing self-hosted bot options, our OpenClaw alternatives roundup includes ZeroClaw alongside nanobot, NanoClaw, memU, and PicoClaw. For a heavier 24/7 assistant, see the OpenClaw setup guide for a heavier 24/7 assistant. If you want something in Python with faster initial setup, the NanoBot setup guide (also updated for GLM-5.3 and MiniMax M3.1) covers that path.
ZeroClaw GitHubWhat this guide covers
- Installing ZeroClaw via the prebuilt install script or Docker
- Configuring MiniMax M3 and GLM-5.3 (Schema V3)
- Web search setup (DuckDuckGo default, Brave upgrade)
- Discord channel integration with correct intents
- Security: risk profiles, gateway pairing, OS sandboxing
- Ops and troubleshooting: verify steps, failure modes, CLI cheat sheet
Video predates v0.8
The video below was recorded against the pre-v0.8 flow. The onboard command and old config format shown are deprecated. Use the written steps in this guide instead.
What’s new in ZeroClaw v0.8 (October 2026)
If you set up ZeroClaw before June 2026, several things you relied on have changed:
Breaking changes if you followed the February 2026 version
Four hard breaks to know about: zeroclaw onboard is deprecated (use zeroclaw quickstart), the config schema moved to V3 (old configs auto-migrate on first load), Docker env vars changed completely (legacy API_KEY/PROVIDER/ZEROCLAW_MODEL are gone), and the gateway port is now 42617 instead of 8080.
onboardreplaced byquickstart- the guided setup now walks through provider, risk profile, memory, channels, and agent in one flow- Schema V3 config - providers, agents, risk profiles, and channels are separate typed sections; old configs auto-migrate, or run
zeroclaw config migrateexplicitly - Prebuilt install script - no Rust toolchain needed; the installer downloads a release binary and falls back to source if needed
- Gateway port 42617 - was 8080; the Docker compose maps
127.0.0.1:42617 - Official Docker images -
ghcr.io/zeroclaw-labs/zeroclawwith distroless and Debian variants; new env-var grammar - Model generation bump - MiniMax M3/M3.1-Flash and GLM-5.3/5.3-Flash replace M2.5 and GLM-5
- Multi-agent runtime - one daemon runs many named agents, each with its own workspace, memory, and provider
- Dashboard, SOP engine, WASM plugins - web UI for config editing and chat, standard-operating-procedure workflows with approval gates, and a plugin host
Old configs are migrated automatically when ZeroClaw loads them. If you want to see what changed before committing, run zeroclaw config migrate and inspect with zeroclaw config list.
What ZeroClaw actually is
ZeroClaw is a self-hosted AI assistant from zeroclaw-labs, built entirely in Rust. The official site claims under 20ms cold start (~12ms measured on Apple Silicon) and memory usage lower than a browser tab. Everything runs from a single config file and the architecture is trait-based: swap your LLM provider, channel, memory backend, or tools by editing TOML. The project is dual-licensed MIT / Apache-2.0 and has 33,000+ stars on GitHub as of October 2026.
The architecture looks like this:
Messages come in from your chat app, hit the gateway (which binds to 127.0.0.1:42617 by default and requires pairing), and the runtime routes them to whatever LLM you configured. The model can use built-in tools (shell, file access, web search, memory) to get work done. Each agent has a risk profile controlling what it can do autonomously.
If you want something even lighter that runs on $10 hardware, PicoClaw, a Go-based assistant that runs on $10 hardware is worth a look.
How it compares
| ZeroClaw | NanoBot | OpenClaw | |
|---|---|---|---|
| Language | Rust | Python | TypeScript |
| Resource footprint | Milliseconds cold start, browser-tab-scale memory | ~100MB RAM | > 1GB RAM |
| Config format | TOML (Schema V3) | JSON | Custom |
| Channels | 30+ | 9 | 4 |
| Providers | 20+ families | 13+ | Several |
| Multi-agent | Yes (v0.8) | No | No |
| Memory | SQLite/Postgres/Qdrant hybrid search | File-based | File + semantic |
| Security | Risk profiles + OS sandbox + pairing + tool receipts | Allowlists | Allowlists |
| Setup | Install script + quickstart | pip install | Custom script |
I stopped printing binary sizes and RAM numbers in this table because they change every release and I can’t re-measure all three tools against their current versions each month. The qualitative story holds: ZeroClaw is the leanest, nanobot is the fastest to set up, and OpenClaw has the most polished OAuth onboarding for existing Claude/ChatGPT subscriptions.
For the nanobot path, see our NanoBot setup guide (also updated for GLM-5.3 and MiniMax M3.1). For OpenClaw, see the OpenClaw setup guide for a heavier 24/7 assistant.
Why MiniMax M3 and GLM-5.3
When I first covered ZeroClaw in February 2026, the budget picks were MiniMax M2.5 and GLM-5. Both are now two generations behind. The current sweet spot for an always-on VPS bot is MiniMax M3.1-Flash or GLM-5.3-Flash - both give you 1M token context windows at prices that make a personal bot affordable.
Pricing as of 2026-10-29 (via OpenRouter; check the model pages before committing):
| Model | Context | Input | Output | Notes |
|---|---|---|---|---|
| MiniMax M3.1-Flash | 1M | $0.15/M | $0.60/M | Flagship-flash, released 2026-10-19 |
| MiniMax M3 | 1M | $0.30/M | $1.20/M | Multimodal (text/image/video-in) |
| GLM-5.3-Flash | 1M | $0.15/M | $0.50/M | Budget pick for agents |
| GLM-5.3 | 1M | check model page | check model page | Reasoning model for SWE/long-horizon tasks |
| GLM-5.3-Prime | 1M | $2.80/M | $8.80/M | High-speed variant, 1.5-2x throughput |
If you want to dig deeper into budget model picks across providers, our cheapest AI models under $1/M tokens guide covers the full list. And if you want to eliminate API costs entirely, you can run local models with Ollama and point ZeroClaw at your local instance.
MiniMax M3.1-Flash
MiniMax M3.1-Flash is the default I’d pick for a Discord bot that handles quick questions and lookups. It has a 1M token context window (so long conversations don’t hit limits), it’s fast, and at $0.15/M input / $0.60/M output it costs less per interaction than most alternatives.
ZeroClaw’s own provider catalog example defaults to MiniMax-M3 with MiniMax-M2.7 and MiniMax-M2.7-highspeed as alternatives. M3.1-Flash is the newer flash-tier option. MiniMax also has a multi-region endpoint field (endpoint = "intl" or "cn") so you can route through the global or mainland China API without hand-building URLs.
GLM-5.3 (Z.AI)
GLM-5.3 from Zhipu AI (Z.AI) is a reasoning model aimed at software engineering and long-horizon agent tasks, released 2026-08-18 with a 1M token context window. It’s a strong pick when your bot does code analysis or multi-step research.
The budget variant GLM-5.3-Flash runs at $0.15/M input / $0.50/M output on OpenRouter. The high-speed GLM-5.3-Prime variant (released 2026-09-23) is $2.80/M input / $8.80/M output with 1.5-2x throughput - overkill for a personal Discord bot but useful if you need low-latency responses under load.
Both models are available through their native APIs and through OpenRouter.
Installation
ZeroClaw now ships a prebuilt install script that handles the binary download - no Rust toolchain required.
Prerequisites
- A Linux VPS (Ubuntu 24.04 works well) or local machine
- A model API key: MiniMax, Z.AI, or OpenRouter
- A Discord account and server (for the Discord integration)
- Port 42617 available (gateway binds to localhost by default)
- No Rust toolchain needed for the binary install path
Install script (recommended)
curl -fsSL https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/master/install.sh | shThe installer downloads a prebuilt release binary (no compile step). If no matching binary exists for your platform, it falls back to building from source and can bootstrap Rust for you.
Then run the guided setup:
zeroclaw quickstartquickstart walks you through provider selection, risk profile, memory backend, channels, and agent creation in one flow. It writes secrets encrypted (API keys go into config.toml with the encryption key in a .secret_key file beside it).
From source
If you want to build from source or need specific channel features:
git clone https://github.com/zeroclaw-labs/zeroclaw.git
cd zeroclaw
./install.sh --sourceThe source build is guided (app/feature picker). For the manual path:
cargo install --locked --path .Channel feature flags: the default build ships a lean bundle: ACP, webhook, email, Telegram, Discord, and filesystem. Release artifacts add Git, Matrix, Lark, and WhatsApp Web. Slack, Mattermost, LINE, Signal, and voice features need --preset full or individual channel-* features. If you add Slack after upgrading and it’s missing, you’re running a lean build.
Verify the install
zeroclaw --version
zeroclaw doctorzeroclaw --version prints the version (Docker images built without the build-arg show “0.8.5 (unknown)” - that’s cosmetic). zeroclaw doctor runs diagnostics and prints warnings about config issues, workspace problems, and channel health.
For a quick smoke test once you have a provider configured:
zeroclaw agent -a assistant -m "hello"If you get a response back, the provider wiring works.
Configuring MiniMax M3
ZeroClaw uses TOML for configuration. The config lives at ~/.zeroclaw/config.toml. Schema V3 organizes providers under typed slots with your own alias.
Get an API key
- Go to platform.minimax.io (global) or minimaxi.com (mainland China)
- Create an account and generate an API key
MiniMax coding plan: 10% off
MiniMax offers coding plans priced for developer workloads. Get 10% off with our referral link. For details on how GLM and MiniMax compare for always-on bots, see our best open source models for OpenClaw breakdown.
Add to config
Three ways to connect, depending on what you have:
OpenRouter
The simplest path if you already have an OpenRouter key - one gateway for any model:
[providers.models.openrouter.default]
model = "minimax/minimax-m3.1-flash"
api_key = "sk-or-your-openrouter-key"Swap minimax/minimax-m3.1-flash for minimax/minimax-m3 if you want the full multimodal model.
MiniMax direct
Use MiniMax’s native typed slot with the multi-region endpoint:
[providers.models.minimax.intl]
model = "MiniMax-M3" # or MiniMax-M2.7 / MiniMax-M2.7-highspeed
api_key = "your-minimax-api-key"
endpoint = "intl" # variants: cn, intlSet endpoint = "cn" for the mainland China API.
Anthropic-compatible (coding plan)
If you’re on MiniMax’s coding plan, use the Anthropic-compatible endpoint:
[providers.models.anthropic.minimax]
model = "MiniMax-M3"
api_key = "your-minimax-api-key"
uri = "https://api.minimax.io/anthropic" # global
# uri = "https://api.minimaxi.com/anthropic" # mainland ChinaTest it
zeroclaw agent -m "What's 42 * 17?"If you get a response, MiniMax is working.
Configuring GLM-5.3 (Z.AI)
GLM-5.3 works through OpenRouter, Z.AI’s native API, or their Anthropic-compatible endpoint.
Get an API key
- Go to z.ai
- Register and create an API key
Z.AI GLM coding plan: 10% off
Z.AI offers GLM coding plans (10% off with our link) designed for continuous developer workloads.
Add to config
OpenRouter
[providers.models.openrouter.glm]
model = "z-ai/glm-5.3-flash" # or z-ai/glm-5.3 for the full model
api_key = "sk-or-your-openrouter-key"Z.AI direct
[providers.models.zai.main]
model = "glm-5.3" # check current model IDs in the Z.AI dashboard
api_key = "your-zai-api-key"Anthropic-compatible (coding plan)
Recommended for coding-plan subscribers:
[providers.models.anthropic.zai]
model = "glm-5.3"
api_key = "your-zai-api-key"
uri = "https://api.z.ai/api/anthropic"Switching between models
Each model lives under its own alias, so switching is editing one line in the [agents.<alias>] section - point model_provider at whichever provider alias you want. No restart needed for the CLI. For the daemon, restart with zeroclaw daemon or zeroclaw service restart.
Web search setup
Without web search, your bot is limited to the model’s training data and whatever files are on your server. The good news: web search now works out of the box with no API key.
No API key needed
DuckDuckGo is the default search provider. It works immediately with zero configuration. Brave, Tavily, SearXNG, and others are optional upgrades.
DuckDuckGo (default)
Search works immediately with no config. If you want to be explicit:
[web_search]
search_provider = "duckduckgo"Brave (upgrade)
Brave Search gives you more control and better results for technical queries. The free tier now gives up to $20 in free calls every month (prepaid credits, no subscription - this changed from the old “1,000 free searches/month” model).
- Go to brave.com/search/api
- Sign up and generate an API key
[web_search]
search_provider = "brave"
brave_api_key = "your-brave-search-api-key"Browser tool settings
The [browser] section controls two different things now:
[browser]
enabled = true # gates browser_open (default true)
automation_enabled = false # gates full browser automation (default false)
allowed_domains = ["docs.rs", "github.com", "stackoverflow.com"]browser.enabled controls the browser_open tool (hands a URL to the system browser). automation_enabled is a separate opt-in for full browser automation - it goes through the approval flow when triggered. The allowed_domains list is shared by both. The default is ["*"]; lock it down once you know what domains your bot needs.
Discord setup
Discord is one of ZeroClaw’s 30+ channels. The integration uses Discord’s WebSocket gateway directly - no webhook URL or public endpoint needed.
Create a Discord bot and enable intents
- Go to discord.com/developers/applications
- Click New Application, give it a name
- Go to Bot in the left sidebar, click Add Bot
- Copy the bot token
Both intents are required
In the Bot settings page, scroll to Privileged Gateway Intents and enable BOTH MESSAGE CONTENT INTENT and SERVER MEMBERS INTENT. Skipping these is the single most common reason a bot sits online but never replies. The old advice that Server Members Intent was optional is wrong - both are required now.
Get your user ID
- Open Discord Settings → Advanced → enable Developer Mode
- Right-click your avatar anywhere in Discord
- Click Copy User ID
Configure ZeroClaw
Add the Discord channel to your config using the Schema V3 shape:
[channels.discord.main]
bot_token = "YOUR_DISCORD_BOT_TOKEN"
allowed_users = ["YOUR_USER_ID"]
# guild_ids = ["YOUR_SERVER_ID"] # scope to specific servers
# channel_ids = ["CHANNEL_ID"] # scope to specific channels
# mention_only = true # only respond when @mentioned
# archive = true # sidecar discord.db + discord_search toolThe bot_token field replaces the old token field. Set it through the encrypted secrets surface (quickstart writes secrets encrypted by default) rather than pasting plaintext into the config file.
Allowlist behavior
An empty allowlist means deny all inbound messages by default - the opposite of most other bots. Add your user ID or use ["*"] for open access. If you’re not sure what your sender identity looks like, start the bot, send it a message, and check the warning log for the exact value.
Useful fields worth knowing: guild_ids and channel_ids scope where the bot listens, mention_only = true makes it respond only when @mentioned (helpful in busy servers), and archive = true stores messages in a local discord.db sidecar with a discord_search tool.
v0.8.1 added slash commands generated from installed skills, and v0.8.2 added approval buttons (interaction components) so you can approve tool actions from Discord. If you’re using the experimental WASM Discord plugin, never give it the same bot token as the built-in channel - you’ll get double replies.
Invite the bot to your server
- In the Discord developer portal, go to OAuth2 → URL Generator
- Under Scopes, check
bot - Under Bot Permissions, check
Send Messages,Read Message History, andView Channels - Copy the generated URL and open it in your browser
- Select the server to add the bot to
Start the daemon and verify
zeroclaw daemonThis starts the full autonomous runtime including all configured channels. Send a message in Discord - the bot should respond.
If nothing happens, run zeroclaw channel doctor for targeted channel health checks. The failure modes table in the Ops section covers the common culprits (intents off, missing permissions, mention_only, invalid token).
Full config example (Schema V3)
If you’re migrating from the old flat config format, ZeroClaw auto-migrates on first load. For an explicit conversion:
zeroclaw config migrate # V2 → V3
zeroclaw config list # print resolved values
zeroclaw config schema # expected shape
zeroclaw config set <key> <value> # set a single keyHere’s what a complete ~/.zeroclaw/config.toml looks like in Schema V3 with OpenRouter, MiniMax M3.1-Flash, Discord, web search, and SQLite memory:
# ── Providers ──────────────────────────────────────────────
[providers.models.openrouter.default]
model = "minimax/minimax-m3.1-flash"
api_key = "sk-or-your-openrouter-key"
temperature = 0.7
# ── Risk profiles ──────────────────────────────────────────
[risk_profiles.standard]
level = "supervised" # readonly | supervised | full
auto_approve = []
always_ask = []
excluded_tools = []
# forbidden_paths = ["/etc", "/root", "/proc", "/sys", "~/.ssh", "~/.aws"]
# ── Agents ─────────────────────────────────────────────────
[agents.assistant]
model_provider = "openrouter.default" # dotted type.alias reference
risk_profile = "standard"
# ── Channels ───────────────────────────────────────────────
[channels.discord.main]
bot_token = "YOUR_DISCORD_BOT_TOKEN" # use encrypted secrets
allowed_users = ["YOUR_USER_ID"]
# ── Web search ─────────────────────────────────────────────
[web_search]
search_provider = "duckduckgo"
# ── Browser ────────────────────────────────────────────────
[browser]
enabled = true
automation_enabled = false
allowed_domains = ["docs.rs", "github.com", "stackoverflow.com"]
# ── Memory ─────────────────────────────────────────────────
[memory]
backend = "sqlite"The full config reference is generated from the live schema at docs.zeroclaw.com.
Provider slots and fallback chains
ZeroClaw has typed provider slots: openrouter, minimax, zai, anthropic, ollama, custom (with uri), and about 20 provider families total. Each slot accepts your alias, so you can run multiple configs for the same provider type.
Since v0.8, you can add fallback chains for resilience:
[providers.models.openrouter.default]
model = "minimax/minimax-m3.1-flash"
api_key = "sk-or-..."
fallback_models = ["z-ai/glm-5.3-flash", "minimax/minimax-m3"]
fallback = "zai.main" # fall back to another provider aliasWorth enabling for an always-on VPS bot: if one provider rate-limits, the request chains to the next.
Memory backends
Memory backends are now sqlite, markdown, postgres, qdrant, lucid, or none. SQLite hybrid search (FTS5 + vectors) remains the default and needs no external services. v0.8.4 added retrieval caching, reranking, and typed-fact extraction.
Set embedding_provider = "noop" if you don’t want to pay for embeddings - you lose vector search but keyword search (FTS5) still works on its own.
Workspace and identity files
ZeroClaw uses workspace files to shape the bot’s personality:
| File | Purpose |
|---|---|
IDENTITY.md |
Who the agent is |
SOUL.md |
Core personality and values |
USER.md |
Your personal info and preferences |
AGENTS.md |
Behavior guidelines |
TOOLS.md |
Tool usage instructions |
BOOTSTRAP.md |
Startup instructions |
MEMORY.md |
Long-term memory notes |
Edit these in ~/.zeroclaw/workspace/.
Bootstrap truncation: with compact_context on (the default), each workspace file is cut at 6,000 characters before reaching the model. Run zeroclaw doctor - it prints [workspace] warnings with retained/discarded counts if your files are getting truncated. Keep personality files concise.
Docker deployment
ZeroClaw now ships official prebuilt images on GitHub Container Registry. The “build it yourself” approach is no longer the primary path.
Official images
| Tag | What it is |
|---|---|
ghcr.io/zeroclaw-labs/zeroclaw:latest |
Distroless (no shell) |
ghcr.io/zeroclaw-labs/zeroclaw:debian |
Debian base with shell, git, curl |
ghcr.io/zeroclaw-labs/zeroclaw:v0.8.7 |
Pinned version (use for prod) |
Multi-arch: amd64 and arm64.
Docker run
docker run -d --name zeroclaw \
-e ZEROCLAW_providers__models__openrouter__default__api_key="$OPENROUTER_API_KEY" \
-e ZEROCLAW_gateway__host=0.0.0.0 \
-e ZEROCLAW_gateway__allow_public_bind=true \
-p 127.0.0.1:42617:42617 \
-v zeroclaw-data:/zeroclaw-data \
ghcr.io/zeroclaw-labs/zeroclaw:latestLegacy env vars are gone
The old API_KEY, PROVIDER, and ZEROCLAW_MODEL env vars were removed in v0.8.0. The only env-var surface is the schema-mirror grammar: ZEROCLAW_<dotted_path_with_double_underscores>=<value>. For example, ZEROCLAW_providers__models__openrouter__default__api_key maps to [providers.models.openrouter.default] api_key.
Docker Compose
version: "3.8"
services:
zeroclaw:
image: ghcr.io/zeroclaw-labs/zeroclaw:latest
container_name: zeroclaw
restart: unless-stopped
ports:
- "127.0.0.1:42617:42617"
volumes:
- zeroclaw-data:/zeroclaw-data
environment:
- ZEROCLAW_providers__models__openrouter__default__api_key=${OPENROUTER_API_KEY}
- ZEROCLAW_gateway__host=0.0.0.0
- ZEROCLAW_gateway__allow_public_bind=true
deploy:
resources:
limits:
cpus: "2"
memory: 512M
healthcheck:
test: ["CMD", "zeroclaw", "status", "--format=exit-code"]
interval: 60s
timeout: 10s
retries: 3
start_period: 10s
volumes:
zeroclaw-data:Download the official compose file (note the branch is master, not main):
curl -O https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/master/docker-compose.ymlARM64 gotcha: the distroless image may exit immediately on ARM64 (Raspberry Pi, Graviton). Use ghcr.io/zeroclaw-labs/zeroclaw:debian instead - it has the shell and libraries needed on ARM.
Hardening notes
- The container runs as non-root (UID 65534) by default.
- Keep the
127.0.0.1:prefix on the port mapping unless you specifically want LAN access. The image setsZEROCLAW_gateway__host=0.0.0.0internally - exposure is bounded by yourports:mapping and firewall. - Pin tags for production (
:v0.8.7) rather than using:latest. - The compose example ships with 2 CPU / 512M resource limits - reasonable for a personal bot.
Security settings
ZeroClaw’s security defaults are stricter than most alternatives.
Gateway pairing
When the gateway starts, it generates a pairing code. You exchange this for a bearer token before any webhook requests are accepted:
# Start gateway (shows pairing code in logs)
zeroclaw gateway
# Exchange code for token
curl -X POST http://127.0.0.1:42617/pair \
-H "X-Pairing-Code: 123456"All subsequent requests need the Authorization: Bearer <token> header. v0.8.5 made pairing codes stronger by default and reading or minting them now requires an owner-only admin token. Auth providers include native (pairing), peercred (Unix socket peer uid), and oidc.<alias> (SSO). A config with require_pairing = false and no other auth is rejected by validation.
Pairing token revocation
To revoke a paired token, use the pairing controls. Editing gateway.paired_tokens in the config does NOT disconnect live connections - already-paired clients stay connected until they’re explicitly revoked.
Risk profiles
Risk profiles replace the old [autonomy] section. Each agent references one:
[risk_profiles.standard]
level = "supervised" # readonly | supervised | full
auto_approve = [] # tools that never prompt
always_ask = [] # tools that always prompt
excluded_tools = [] # tools that are disabledreadonly- no writes, no shell execution.supervised(default) - medium-risk tools prompt through the initiating channel (Discord shows an approval button). The defaultapproval_timeout_secsis 120; a timeout is treated as denial.full- autonomous execution. The OS sandbox still applies.
Per-agent tool allowlists are enforced at every dispatch path.
OS sandboxing
[risk_profiles.standard]
sandbox_enabled = true
sandbox_backend = "landlock" # landlock | bubblewrap | seatbelt | docker
forbidden_paths = ["/etc", "/root", "/proc", "/sys", "~/.ssh", "~/.gnupg", "~/.aws"]For the sandboxing context and how it compares to other tools, see our OpenClaw security guide covering CVE-2026-25253.
What’s new in security since February
v0.8 added several security layers worth knowing about: tool receipts (cryptographic proof on every action - a claimed-but-not-made action produces a missing/invalid receipt), a universal ingress policy layer (v0.8.2), SSRF guards on outbound HTTP tools, and private-host allowlists. The allow_public_bind setting silences the startup warning when binding to 0.0.0.0 - it does not itself refuse the bind. Your firewall and port mapping are the real gate.
Ops and troubleshooting
This is the section the February version of this guide lacked. Verify first, then debug.
Verify checklist
zeroclaw –version- confirms the binary is installedzeroclaw doctor- full diagnostics (config, workspace, channels)zeroclaw channel doctor- targeted channel health checkszeroclaw status- runtime status summaryzeroclaw config list- print resolved config valueszeroclaw agent -a assistant -m “hello”- smoke test with your provider
Failure modes
| Symptom | Likely cause | Fix |
|---|---|---|
| Bot online but never replies | Intents not enabled | Enable Message Content AND Server Members intents in the Discord portal |
| Bot replies nowhere | Not invited or missing permissions | Re-invite with Send Messages + Read Message History + View Channels |
| Bot ignores most messages | mention_only = true |
Set to false or @mention the bot |
| Auth errors on startup | Invalid bot token | Reset the token in the Discord portal and update config |
| “Provider streaming failed, falling back to non-streaming chat” | Provider streaming issue | Run zeroclaw auth status to check credentials |
| Container exits immediately (ARM64) | Distroless image on ARM | Use :debian tag |
| Slack/Mattermost/LINE missing after upgrade | Lean build | Rebuild with --preset full or channels-full |
| Config errors on startup | Old schema | Run zeroclaw config migrate |
Diagnostics
zeroclaw service logs --follow # tail service logs
RUST_LOG=debug zeroclaw daemon # verbose debug output
zeroclaw self-test # self-check
zeroclaw estop # emergency stopUpgrades
zeroclaw updateThe dashboard (v0.8.4+) also supports in-app upgrades with automatic restart.
CLI cheat sheet
| Command | What it does |
|---|---|
zeroclaw quickstart |
Guided setup (replaces onboard) |
zeroclaw agent -m "..." |
Send a single message |
zeroclaw agent |
Interactive chat mode |
zeroclaw daemon |
Start full autonomous runtime |
zeroclaw gateway |
Start webhook server |
zeroclaw status |
Show system status |
zeroclaw doctor |
Run diagnostics |
zeroclaw channel doctor |
Check channel health |
zeroclaw config list |
Print resolved config |
zeroclaw config schema |
Show config schema |
zeroclaw config set <key> <value> |
Set a config key |
zeroclaw config migrate |
Migrate old config to V3 |
zeroclaw auth login |
Authenticate with a provider |
zeroclaw auth status |
Check auth credentials |
zeroclaw memory |
Memory management |
zeroclaw models |
List available models |
zeroclaw providers |
List provider slots |
zeroclaw agents |
List configured agents |
zeroclaw skills |
Manage skills |
zeroclaw cron |
Scheduled tasks |
zeroclaw sop |
SOP engine |
zeroclaw update |
Update ZeroClaw |
zeroclaw estop |
Emergency stop |
zeroclaw self-test |
Self-check |
zeroclaw service install |
Install as system service |
zeroclaw service start |
Start the service |
zeroclaw service status |
Check service status |
zeroclaw service logs |
View service logs |
zeroclaw migrate openclaw |
Migrate memory from OpenClaw |
The migrate openclaw command imports memory from an OpenClaw installation. Run --dry-run first to preview what gets imported, then --reindex to rebuild the search index after migration. It migrates memory only - not channels or config.
Impersonation warning
zeroclaw-labs/zeroclaw is the only official repo. The self-hosted bot space has a lot of forks and clones. If a repo doesn’t match that exact path, it’s not the official ZeroClaw.
ZeroClaw vs nanobot vs OpenClaw
Where each one stands as of October 2026:
| Aspect | ZeroClaw | NanoBot | OpenClaw |
|---|---|---|---|
| Language | Rust | Python | TypeScript |
| Config format | TOML (V3) | JSON | Custom |
| Resource footprint | Milliseconds cold start, browser-tab memory | ~100MB RAM | > 1GB RAM |
| Channels | 30+ | 9 | 4 |
| Providers | 20+ families | 13+ | Several |
| Multi-agent | Yes (v0.8) | No | No |
| Memory | SQLite/Postgres/Qdrant hybrid | File-based | File + semantic |
| Security | Risk profiles + OS sandbox + pairing + tool receipts | Allowlists | Allowlists |
| Dashboard | Yes (v0.8) | No | Yes |
| Setup | Install script + quickstart | pip install | Custom script |
ZeroClaw wins on resource usage, security depth, and channel count. nanobot wins on initial setup speed. OpenClaw has the most polished OAuth onboarding for existing Claude/ChatGPT subscriptions and a dashboard.
For the nanobot path, see the NanoBot setup guide (also updated for GLM-5.3 and MiniMax M3.1). For OpenClaw, see the OpenClaw setup guide for a heavier 24/7 assistant. For a Go-based alternative on cheap hardware, check out PicoClaw, a Go-based assistant that runs on $10 hardware. Our OpenClaw alternatives roundup covers the full list.
Frequently asked questions
How much does it cost to run ZeroClaw?
VPS: roughly $5-7/month (Hetzner CX22 or equivalent). API costs depend on usage, but with M3.1-Flash or GLM-5.3-Flash at $0.15/M input, a personal Discord bot typically runs $5-20/month total. ZeroClaw has a built-in usage ledger with an offline pricing catalog (v0.8.3) so you can track spend without guessing.
Can I use ZeroClaw without any API costs?
Yes. Set up an Ollama provider slot and point it at a local Ollama instance. You need hardware that can run inference, but there are no API bills. See run local models with Ollama for the setup.
Does ZeroClaw work on a Raspberry Pi?
Yes. Prebuilt arm64 binaries are available through the install script, so you don’t need to compile from source. The binary install path avoids the OOM issues that plagued source builds on 1GB Pi boards. The Debian Docker image works on ARM64 if you prefer containers.
Can multiple people use one ZeroClaw instance?
Yes. Add multiple user IDs to the channel allowlist. Each person gets their own conversation context. Since v0.8, you can also run multiple named agents - each with its own workspace, memory, and provider - from a single daemon.
What’s the difference between gateway and daemon?
zeroclaw daemon runs the full autonomous runtime: channels, agents, scheduling, heartbeats - everything. zeroclaw gateway starts just the webhook/API server. For 24/7 Discord use, run the daemon (or install it as a service with zeroclaw service install).
Can I migrate from OpenClaw to ZeroClaw?
Yes. Run zeroclaw migrate openclaw --dry-run first to preview what gets imported, then run it for real. Add --reindex to rebuild the search index after migration. Note: it migrates memory only, not channels or config.
Can I add Telegram alongside Discord?
Yes. Configure both channels in the same config file under their respective [channels.<type>.<alias>] blocks. ZeroClaw handles them simultaneously. See the nanobot guide’s Telegram section for the BotFather flow since the process is the same.
What happened to zeroclaw onboard?
It’s deprecated as of v0.8. The command prints a deprecation notice and its flags no longer apply. Use zeroclaw quickstart instead - it walks through provider, risk profile, memory, channels, and agent creation in one guided flow.
VPS hosting and monthly cost
ZeroClaw runs on modest hardware. A 2 vCPU / 4GB VPS is more than enough - the bot’s own footprint is tiny and the real resource consumer is whatever Docker containers you’re already running alongside it.
Monthly cost breakdown (October 2026):
- VPS: ~$5-7/month (Hetzner CX22 or equivalent)
- API: $5-20/month for personal use with M3.1-Flash or GLM-5.3-Flash
- Search: $0 (DuckDuckGo default) or up to $20/month in free Brave credits
- Local models: $0/month if you use Ollama instead of API providers
The built-in usage ledger (v0.8.3) tracks token spend against an offline pricing catalog, so you can see exactly what each conversation costs.
Hetzner discount
Get €20 credit or try Hostinger VPS when you sign up through our referral links. That covers several months of a small VPS. For other providers, Vultr and DigitalOcean work fine too.
For more on how VPS pricing shifted this year, see the Hetzner Cloud review after the 2026 price changes and what changed and when a mini PC wins.
Quick setup on a fresh Ubuntu 24.04 VPS:
ssh root@YOUR_SERVER_IP
# Update system
apt update && apt upgrade -y
# Install ZeroClaw (no Rust needed)
curl -fsSL https://raw.githubusercontent.com/zeroclaw-labs/zeroclaw/master/install.sh | sh
# Guided setup
zeroclaw quickstart
# Verify
zeroclaw doctor
# Install as a systemd service
zeroclaw service install
zeroclaw service start
zeroclaw service statusThis creates a systemd service that starts on boot and restarts on failure. Check logs with zeroclaw service logs --follow.
If you want to run local models alongside ZeroClaw, check our guide on installing Ollama with Docker. ZeroClaw’s Ollama provider connects to any local instance without extra config.
If you want to explore other AI coding tools, our AI coding tools comparison covers the current options. For a self-improving assistant with voice mode and OpenClaw migration, see the Hermes Agent setup guide. For MCP basics that work across all these assistants, check the MCP introduction for beginners.
Did this guide work for you?
Need help with Astro, VPS, or a self-hosted app? View services →


