[ TRANSMISSION SIGNED BY PAPERCHASELABS ]
FREQ 140.85 · CH 401 · CLI
paperchaselabs.com

The Sovereign Operator System. Autonomous AI operator CLI — Claude-Code-style REPL meets Hermes-class autonomous loop. Memory engine, multi-agent dispatch, MCP server, skill registry. Operator-owned. Sovereign by default. MIT.

pip install paperchase
paperchase

Built by PaperChaseLabs.

# canonical
pip install paperchase

# zero-install one-shot
uvx paperchase

# js world
npx paperchase

# macOS one-line
brew install chasewebb/paperchase/paperchase

# Interactive REPL with Ollama (free, local). Codec banner on boot.
paperchase chat

# Autonomous loop — PLAN → ACT → CRITIQUE, with subagent fan-out
paperchase auto "summarize the readme of https://github.com/openai/openai-python"

# Run as an MCP server (Claude Desktop, Cursor, any MCP client)
paperchase serve --mcp-stdio

# Install skills — three sources supported
paperchase skill install builtin:web-research
paperchase skill install github:operator/some-skill
paperchase skill install ./my-local-skill

# Reflect on recent sessions — writes ~/.paperchase/learnings.md
# The operator persona auto-loads it on every future session
paperchase reflect

# Runtimes, vault, skills, lab links
paperchase status

The only autonomous operator CLI that's MIT, local-first, ships an MCP server, and supports every package manager you already use.

• Interactive REPL with codec UI (MGS1 codec aesthetic, animated boot)
• Autonomous loop — PLAN/ACT/CRITIQUE state machine with halt conditions
• SpawnAgent tool — child loops for parallel work, budget-guarded
• 9 tools: Read · Write · Edit · Glob · Grep · gated Bash · WebFetch (HTTPS only) · Skills · SpawnAgent
• Skill registry: install from builtin:, local path, or github:owner/repo[@ref]
• Self-reflection: paperchase reflect reads vault, distills patterns, writes ~/.paperchase/learnings.md — auto-loaded into the persona prompt every future session
• Vault: SQLite at ~/.paperchase/vault.db, every session/turn/tool-call recorded, sessions resumable
• 3 runtimes: Ollama (default, local, free), Anthropic (opt-in via env), OpenAI (opt-in via env)
• MCP server: stdio + HTTP transports, exports all 9 tools
• Polyglot install: pip · uvx · npx · brew

PaperChase ships with one reference skill (web-research). Skills are pluggable abilities the agent can absorb and switch between like modes. Install from anywhere:

paperchase skill install builtin:web-research
paperchase skill install github:owner/repo
paperchase skill install ./my-local-skill

Each skill is a directory with a skill.toml manifest + a Python entry point. See docs/SKILLS.md.

Drop this into ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "paperchase": {
      "command": "paperchase",
      "args": ["serve", "--mcp-stdio"]
    }
  }
}

Zero telemetry. Zero data collection. No usage pings, no error reports, no metrics shipped anywhere. Your vault is a local SQLite file. Ollama runs on your machine. Anthropic/OpenAI runtimes are opt-in via env vars — your keys, your bill, your data. WebFetch goes only where the operator (or agent) points it. Skills installed from GitHub clone over public HTTPS, no auth.

The only network call PaperChase initiates without you asking is the runtime call to Ollama at localhost:11434. That's it.

• Skill marketplace (curated registry + UI)
• Voice in/out (Whisper input + Grok voice output)
• Channel adapters: Telegram, Discord, Slack — agent reachable from anywhere
• Self-fine-tune loop (capture preferred trajectories, fine-tune nightly)
• Embeddings-based vault retrieval (v0.1 uses FTS5)

• Lab — https://paperchaselabs.com
• Discord — https://discord.gg/5vYFCcKVrB
• Substack — https://substack.com/@paperchase
• Book a call — https://calendly.com/paperchasewebb/15min

_{PAPERCHASEWEBB INC. · BUILT IN HONOLULU · 2018→}