acatome-chat

A local-first AI research assistant for scientific literature.

Runs with a local Ollama model by default — or switch to Claude, GPT-4o, or any litellm-compatible provider with a single flag. One install gives you an interactive shell with a searchable paper library, document editing, live web search, and domain-specific databases — all wired together through the Model Context Protocol (MCP).

What you get

Install

pip install acatome-chat
# or
uv add acatome-chat

That's it. All MCP servers and the paper store are included. Default backend is SQLite + Chroma — no external services needed.

For heavier setups:

uv add "acatome-chat[postgres]"     # PostgreSQL + pgvector
uv add "acatome-chat[embeddings]"   # sentence-transformers

Quick start

1. Build your paper library

Extract PDFs and ingest them into the searchable store:

# Extract a single PDF (or a whole directory)
acatome-extract extract paper.pdf
acatome-extract extract ~/Downloads/papers/

# Ingest extracted bundles into the searchable store
acatome-store ingest ~/.acatome/papers/

Optional enrichment steps:

# Watch a folder for new PDFs (auto-extracts on arrival)
acatome-extract watch ~/Downloads/papers/

# Add LLM-generated summaries to your bundles
acatome-extract enrich ~/.acatome/papers/

2. Start the chat

# Default: local Ollama model (ollama/qwen3.5:9b)
acatome-chat

# Or use Claude / GPT-4o / any litellm provider
acatome-chat --model anthropic/claude-sonnet-4-20250514
acatome-chat --model openai/gpt-4o
acatome-chat --model ollama/llama3.1:8b

# Disable thinking/reasoning mode
acatome-chat --no-think

3. Use slash commands

The shell has / commands with tab autocomplete:

4. Review documents

› /review check scientific rigor and paragraph structure
› /review fix grammar and clarity --fix
› /review verify citations --comments-only
› /review improve transitions --section S2.3

The assistant reads each section, applies fixes as tracked changes, and adds margin comments for issues needing human judgment.

5. Talk to your papers

After ingesting papers, the LLM has direct access to your library. Example prompts:

› Find papers about CO2 conversion and write a summary with citations into co2review.docx
› Search for MOFs with high CO2 uptake and compare their pore sizes
› Read the abstract of li2024mof and summarize the key findings
› Open my draft.docx and add a new section about these results
› Search the web for recent advances in direct air capture

The assistant can read your papers, search the web, query chemistry databases, and write results directly into .docx or .tex files — all in one conversation.

Note: Document editing supports .docx and .tex formats. For Word files, changes are written as tracked changes. Be aware that Word may overwrite the file if it's open — save and close Word before asking the assistant to edit.

Architecture

acatome-chat (you are here)
├── acatome-lambic        LLM shell engine (MCP client)
├── acatome-mcp           Paper query MCP server
│   └── acatome-store     SQLite/Postgres + Chroma/pgvector storage
│       └── acatome-meta  Shared config and metadata
├── acatome-extract       PDF → structured bundle pipeline
│   └── precis-summary    RAKE keyword extraction
├── precis-mcp            Document editor MCP server
├── perplexity-sonar-mcp  Web search MCP server
├── catapult-mcp          Catalysis database MCP server
│   └── chemdb-common     Shared chemistry DB utilities
└── grandmofty-mcp        MOF database MCP server
    └── chemdb-common

Environment variables

For local models via Ollama, no API keys are needed.

License

GPL-3.0-or-later