2026-05-10
Picobot v1 — 9MB Go binary lands as the minimum-viable production agent
Picobot ships as a single 9MB Go binary, 10MB idle RAM, 29MB Alpine image. 16 built-in tools, OpenAI-compatible. Runs on $5 VPS, Pi, old Android.
Picobot v1 launched on May 9, 2026 — a single ~9MB compiled Go binary with ~10MB idle RAM and a 29MB Alpine Docker image. It’s the strongest argument yet that “production agent” doesn’t require Python + Docker Compose + a vector DB sidecar.
What’s in the box
16 built-in tools out of the box:
| Category | Tools |
|---|---|
| Filesystem | read, write, list, search, delete |
| Execution | shell exec, cron schedule |
| Web | fetch URL, web search |
| Memory | persistent KV CRUD |
| Orchestration | spawn sub-agent, manage skills |
API surface: any OpenAI-compatible endpoint — OpenAI itself, OpenRouter, Ollama, vLLM, llama.cpp server, LM Studio. The same binary talks to all of them via env vars.
Chat integrations: Telegram, Discord, Slack, WhatsApp out of the box. Each is a single-flag toggle.
Demonstrated targets: $5/month VPS, Raspberry Pi 4, an old Android phone running Termux. Idle RAM stays at ~10MB. Active RAM peaks at ~50MB during heavy tool use.
Why size matters here
The standard “agent in production” story today involves Python with 100+ MB of dependencies, a vector DB process, often a separate orchestration service, and a Docker image typically 800 MB - 2 GB. That’s fine for a SaaS data center. It’s miserable for:
- IoT / edge agents that must run on cheap ARM SBCs
- Side-project bots where a 9MB binary saves you Docker Compose entirely
- Auxiliary fleet workers (e.g., one agent per server doing log triage) where the per-instance footprint determines fleet density
Picobot’s 9MB also means you can scp it to a target and run, no toolchain required.
Where Picobot stops
It’s not a replacement for Claude Code or Cursor. Specifically:
- No IDE integration
- No code-completion model
- No multi-file refactoring intelligence beyond what the underlying LLM provides
- No persistent context across machines (memory is local-disk KV, not networked)
For “small recurring task on cheap hardware” it’s perfect. For “I’m pair-programming all day,” still use the heavier tools.
Practitioner note
If you’ve been carrying around a Python agent stack just because that’s where the libraries are, Picobot is worth a 30-minute spike. Concretely: pick one cron-style task you’re already running (RSS digest, log triage, daily report email) and rebuild it as a Picobot config. If the experience is good, you’ve found a new default for low-stakes recurring work. The 16-tool floor is also useful as a reference set — when designing your own agents, ask “would I add a tool that’s not in this 16?” and if no, you have a tight, well-named tool surface to copy.