NanoClaw

Made by: Community (Zig) Type: Ultra-lightweight Agent Runtime Best for: Edge deployments, embedded systems, serverless, anywhere size and startup time matter

What it is

NanoClaw is an agent runtime built in Zig — a systems programming language optimised for performance and binary size. The result is remarkable:

• 678KB binary — smaller than most favicons, smaller than a JPEG
• < 2ms startup time — faster than a blink
• 50+ provider integrations — Telegram, Slack, Discord, email, REST APIs, and more
• 19 channel types — every major messaging platform
• Zero runtime dependencies — the binary is the entire runtime

It's OpenClaw's smaller, faster sibling. Where OpenClaw prioritises features and extensibility, NanoClaw prioritises footprint and speed.

Why size and startup time matter

Most agents run in the cloud where resources are plentiful. NanoClaw targets the cases where that's not true:

Serverless functions — Cold starts are the enemy. A 678KB binary starting in 2ms vs a Node.js runtime taking 800ms is a meaningful difference at scale.

Edge computing — Running agents close to users (Cloudflare Workers, Fastly Compute) requires extreme size constraints.

Embedded systems — IoT devices, industrial controllers, smart appliances — these can now run agents.

High-concurrency deployments — When you're running thousands of agent instances simultaneously, footprint per instance adds up fast.

Kubernetes pods — A pod image with NanoClaw can be 5MB vs 500MB for a Node.js equivalent. Faster pod startup, lower registry costs, faster scaling.

Channels supported

NanoClaw has native support for 19 channels including:

Configuration

NanoClaw uses a TOML config file:

[agent]
name = "my-nanoclaw-agent"
model = "claude-sonnet-4-5"
system_prompt = "You are a helpful assistant."
 
[channels.telegram]
bot_token = "${TELEGRAM_BOT_TOKEN}"
 
[channels.slack]
token = "${SLACK_BOT_TOKEN}"
 
[memory]
type = "sqlite"
path = "./memory.nc"
 
[schedule]
cron = "0 9 * * *"
task = "Send daily briefing to Telegram"

Memory

NanoClaw uses SQLite with FTS5 for agent memory — the same search primitives that power enterprise search engines. Agents can store and retrieve information with full-text search, ranking, and fuzzy matching — all in a 500KB database file with no server required.

[memory]
type = "sqlite"
path = "./agent.db"
max_entries = 10000
search = "bm25"   # BM25 relevance ranking

Sub-agents (delegate tool)

NanoClaw supports sub-agent delegation — an agent can spin up another NanoClaw instance, give it a task, and incorporate the result:

[[tools]]
type = "delegate"
agent = "research-agent"
config = "./research-agent.toml"

This enables multi-agent patterns even in resource-constrained environments. Because each NanoClaw instance starts in under 2ms, spawning a sub-agent has negligible overhead.

Scheduling

Built-in cron scheduling, no external scheduler needed:

[[schedules]]
cron = "*/5 * * * *"
task = "Check monitoring alerts"
 
[[schedules]]
cron = "0 8 * * 1-5"
task = "Send morning standup summary to Slack"

Deployment

# Download the binary (one file, 678KB)
curl -fsSL https://nanoclaw.dev/install | bash
 
# Run with a config file
nanoclaw --config agent.toml
 
# Or pipe config via stdin
cat agent.toml | nanoclaw
 
# Run in the background (systemd, supervisor, or just &)
nanoclaw --config agent.toml &

Docker:

FROM scratch
COPY nanoclaw /nanoclaw
COPY agent.toml /agent.toml
ENTRYPOINT ["/nanoclaw", "--config", "/agent.toml"]

Final image: ~1MB.

NanoClaw vs OpenClaw

When NanoClaw makes sense

• You're running hundreds or thousands of agent instances
• You need agents on serverless (Lambda, Cloud Functions, Workers)
• You're deploying to IoT or embedded hardware
• You need the fastest possible response time to user messages
• You want the simplest possible deployment (one file, no dependencies)
• You're building an agent that runs as a Telegram or Discord bot and needs to be always-on with minimal cost

Resources

• GitHub: github.com/nanoclaw/nanoclaw (opens in a new tab)
• Docs: nanoclaw.dev/docs (opens in a new tab)