ZIP-0807: zoo-bot

Abstract

zoo-bot is a pure-Go re-write of Hanzo's OpenClaw agentic bot framework (~/work/hanzo/bot). It preserves the OpenClaw skill manifest format and MCP wire protocol so skills written for OpenClaw run on zoo-bot via a compatibility shim, while replacing the TypeScript/Node runtime with a single static Go binary. Goals: smaller binary (< 50 MB), lower idle memory (< 80 MB), faster cold start (< 100 ms), narrower attack surface (no npm dependency tree).

Motivation

OpenClaw is good. Its Node runtime is appropriate for front-end-style workloads, less appropriate for continuous-loop server agents:

npm dependency tree exposes transitive supply chain risk.
Node runtime + TypeScript JIT carry several hundred MB of overhead.
Cold start is multi-second.
Sandboxing is cooperative.

A Go re-write addresses all four with a different tradeoff: less dynamism (no runtime npm install of tools) in exchange for the operational properties above.

Specification

Runtime

Single static binary (Go 1.x).
SQLite or Postgres persistence; SQLite by default for local dev.
Embedded MCP server and client.

Skill interface

type Skill interface {
    Name() string
    Plan(ctx context.Context, goal string) (*Plan, error)
    Execute(ctx context.Context, plan *Plan) (*Result, error)
    Observe(ctx context.Context, result *Result) error
}

Each skill is a Go module satisfying this interface. Skills are compiled into the binary (no runtime plugin loading) for reproducibility and supply-chain safety.

OpenClaw compatibility shim

A compatibility layer reads OpenClaw skill manifests (skill.json) and translates them to Go skill descriptors. Skills declared in TypeScript run via an embedded V8 isolate as a fallback path; new skills should be authored in Go directly.

Adapters

Discord, Slack, Matrix, generic webhook.
Zoo L1 RPC adapter for on-chain trigger streams.
Lux RPC adapter for cross-chain triggers.
A-Chain attestation feed adapter.

Sandbox

Each skill runs in a Goroutine with capability-bounded enforcement:

Linux: seccomp-bpf profile.
macOS: sandbox_init() profile.
Windows: AppContainer integrity level.

Capabilities declared in the skill manifest:

{
  "name": "leaderboard-poster",
  "filesystem": ["/var/zoo-bot/cache"],
  "network": ["api.discord.com:443"],
  "rpc": ["lux:zoo-l1:rpc"]
}

A skill cannot escape its declared capabilities without source modification + re-deploy.

Use cases

Agentic workflows (per-LLM chain leaderboards, attribution graph visualisations).
Content automation (A-Chain provenance feed publication).
On-chain triggers (NFT price alerts, training receipt notifications).
Validator-side automation (mempool anomaly paging).

Rationale

A re-write rather than a fork keeps the OpenClaw upstream maintainable for Hanzo's TypeScript ecosystem. The shared MCP wire protocol means a zoo-bot agent can call OpenClaw tools and vice versa; no ecosystem fragmentation.

Reference Implementation

~/work/zoo/bot (planned). Existing Hanzo OpenClaw at ~/work/hanzo/bot.

Security Considerations

Skills are statically compiled into the binary; no runtime code load reduces dynamic supply-chain risk.
The OS-level sandbox is not optional; skills cannot be deployed without a manifest declaring capabilities.
Persistence is encrypted at rest.

References

zoo-3-0-launch paper §6 (zoo-bot)
Hanzo OpenClaw: ~/work/hanzo/bot
MCP specification (~/work/hanzo/mcp)

zoo-bot — Pure-Go Agentic Bot Framework