PicoFlare

Cloudflare-native AI agent with Code Mode MCP, R2 storage, and Vectorize memory.

What makes PicoFlare different: LLM-defined agent scaling

Describe an agent, get it. PicoFlare scales agents differently: instead of hand-coding each specialist or wiring up complex orchestration, you tell the LLM what kind of agent you want and it creates it for you.

• /createagent – Describe an agent in plain language (e.g. "Next.js specialist for App Router" or "Python data science expert for pandas"). The LLM writes a skill to workspace/skills/<name>/SKILL.md and it’s loaded into context for future use.
• Agent self-creation – The agent itself can create new agents. Ask it to "create a Rust async specialist" and it uses the create_skill tool to add that capability. No code changes, no redeploys.
• Skills as domain knowledge – Each skill is markdown with YAML frontmatter, injected into the system prompt. New agents become first-class capabilities the main agent can draw on.

This is the core differentiator: agent scaling via natural language, not pipelines or hardcoded roles.

What it does

• Cloudflare MCP – Calls the Cloudflare MCP via Streamable HTTP (JSON-RPC). Two tools: search (query the OpenAPI spec) and execute (run API calls via Code Mode). ~1k tokens instead of ~244k.
• R2 Storage – S3-compatible object storage via Cloudflare R2.
• Vectorize – Vector memory for RAG (semantic search over stored knowledge).
• Telegram – Bot channel with /createagent and voice notes.

Setup

cp .env.example .env
# Fill in your Cloudflare credentials and Telegram token

Build & Run

go build -o picoflare .
./picoflare              # default: run pico-flare agent (interactive)
./picoflare agent        # run pico-flare agent (interactive)
./picoflare bot          # Telegram bot (TELEGRAM_BOT_TOKEN required)
./picoflare mcp-test     # create R2 bucket + Vectorize index via MCP
./picoflare help         # show usage

When the MCP server is unavailable, pico-flare agent falls back to the Cloudflare REST API so you still get Workers, R2, KV, D1, and Vectorize tools.

Webhook mode (Cloudflare deployment)

Set TELEGRAM_WEBHOOK_URL to run behind Cloudflare Tunnel or a reverse proxy:

TELEGRAM_WEBHOOK_URL=https://picoflare.example.com/bot ./picoflare bot

See DEPLOY_CLOUDFLARE.md for full instructions.

Project Structure

PicoFlare/
├── main.go                  # Entry point, bot / mcp-test commands
├── pkg/
│   ├── agent/               # Agent loop, Code Mode tools, create_skill
│   ├── bot/                 # Telegram handlers, /createagent
│   ├── mcpclient/client.go  # Cloudflare MCP client (Streamable HTTP)
│   ├── skills/loader.go     # Load SKILL.md from workspace/skills/*
│   ├── storage/r2.go        # R2 object storage (S3-compatible)
│   └── memory/vectorize.go  # Vectorize REST client (RAG memory)
├── skills/
│   └── mcp-builder/         # Built-in skill: create MCP servers on Workers
├── workspace/skills/        # LLM-created agents (via /createagent)
├── AGENT_DESIGN.md          # Architecture + token-aware design
├── .env.example             # Template (no secrets)
└── .gitignore

Design

See AGENT_DESIGN.md for the full architecture: MCP Builder skill, token tracking, and optimization principles. The agent scaling model (LLM-defined skills via /createagent and create_skill) is documented there as well.

License

MIT