scion 0.2.1

Turn a coding-AI chat session into a deployable, continuously-learning SME agent served over MCP.

scion

Turn a coding-AI chat session (Claude Code or Codex CLI) into a deployable, continuously-learning subject-matter-expert agent that any MCP-aware host can talk to.

scion extracts everything from a Claude Code or Codex CLI session — every message, every tool call, every memory file, every skill, every config — distills it into a portable, git-committable Agent Bundle, and serves the bundle as a stdio MCP server. Your colleague's Claude Code or Codex CLI calls /mcp__scion-<name>__ask (or the query tool) and gets a grounded answer. The bundle learns from every interaction via a nightly dream loop that consolidates new episodes into memory, gated by an evaluation ratchet so the agent cannot silently degrade.

Status

Pre-alpha. Building toward v0.1.0 (Phases 1–4 of the PRD).

Phase	What	State
0	Foundation, governance, context docs	in progress
1	Claude Code extractor + bundle + CLI + LLMClient skeleton	planned
2	Stdio FastMCP server + distill + memory tool wiring	planned
3	Episode writeback + nightly dream loop with eval ratchet	planned
4	Codex CLI extractor (closes v0.1.0)	planned

Three-command quickstart (target — not yet runnable)

uv tool install scion
scion init demo
scion demo                  # full pipeline against synthetic fixture, no API key

Once Phase 1 ships, replace the third line with:

scion extract claude-code --cwd ~/proj/recon --output ./bundles/recon-sme
scion install claude-code ./bundles/recon-sme
# Restart Claude Code → /mcp__scion-recon-sme__ask "how do partial refunds work?"

How it compares

	scion	mem0	Letta / MemGPT	Zep CE
Local-first, no daemon	✓	hosted	daemon	needs Neo4j
Git-committable bundles	✓	✗	✗	✗
Continuous learning loop	eval ratchet (Karpathy)	LLM-extracted facts	tiered memory (managed)	temporal graph
Memory shape	files (markdown)	facts (KV+vector)	tiered store	knowledge graph
Provider portability	LiteLLM (any)	OpenAI-default	OpenAI-default	provider-agnostic
Plugin extractor surface	✓	n/a	n/a	n/a
Source-tool side effects	none (read-only)	n/a	n/a	n/a

Supported on

Linux + macOS, Python 3.11/3.12/3.13. Windows is out of scope for v0.1.0.

License

Apache-2.0.

Contributing

See CONTRIBUTING.md. DCO sign-off required (git commit -s); a hook is provided.