langchain-curt

LangChain chat loader for .curt — the open (MIT), single-file AI conversation format. One timestamped, attributed conversation per file; unknown fields are data and survive round-trips.

from langchain_curt import CurtChatLoader

loader = CurtChatLoader("./conversations/*.curt")
sessions = loader.lazy_load()   # one ChatSession per .curt file

• Role mapping: user → HumanMessage, assistant → AIMessage, system → SystemMessage. tool messages are skipped and counted (loader.stats["tool_messages_skipped"]) — ToolMessage requires a tool_call_id that .curt v1 does not carry.
• Metadata: per-message timestamps (epoch ms → ISO 8601 UTC), model, and all unknown/extension fields under additional_kwargs["curt"]; conversation-level metadata in loader.conversation_metadata.
• Validation: spec-compliant whole-file rejection (bad magic, wrong version, missing fields → rejected, never half-loaded). on_error="skip" imports valid files and reports invalid ones in loader.errors.
• Dependencies: langchain-core only. The .curt parser is stdlib-only.

Converters into .curt exist for Claude Code transcripts (curt-convert), Open WebUI exports (curt-openwebui), and the WildChat-1M research corpus (curt-wildchat).

Spec: https://github.com/JordanNewell/dot_curt · npm reference impl: curt-format

Development

pip install -e . pytest
pytest

MIT License.