Thoth Research Assistant
A local-first, extensible research assistant for staying on top of AI/ML papers. Built with Python, Letta, and Obsidian.
接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
來自發布者
A local-first, extensible research assistant for staying on top of AI/ML papers. Built with Python, Letta, and Obsidian.
An advanced AI-powered research assistant system built for flexibility, extensibility, and user control. Thoth combines intelligent automation with comprehensive customization — every prompt, schema, and source can be tailored to your workflow.
As an AI practitioner juggling a full-time job while trying to stay current with rapid advancements in AI research, I found myself drowning in papers. Existing research tools were too rigid—they locked me into their workflow, their sources, their extraction schemas. I couldn't easily add new paper sources, customize what metadata got extracted, or control where my data lived.
But the real friction wasn't just the lack of customization—it was the interface. Research tools forced me to constantly hunt through settings files, create countless memory configurations, edit YAML configs, and manage dozens of rule files. Sure, I could edit these manually, but why should that be the only option?
Users want a single integration point: natural language. They don't want to manage files—they want to tell the agent what they need and have it happen.
This drove a core architectural decision: the agent has access to everything a user can do manually. Every setting, every prompt template, every research query, every extraction schema—the agent can read and modify all of it through MCP tools. The system is deeply configurable, but users shouldn't have to learn the configuration surface to use it.
In practice, this means you can say things like:
Advanced users can still edit settings.json, prompt templates, and schemas by hand. But most of the time, you just talk to the agent. There's also a TUI setup wizard (thoth setup) for the initial configuration, so the first-run experience isn't overwhelming either.
After trying various agent frameworks, Letta (formerly MemGPT) was the first that actually remembered what mattered. Its self-editing memory architecture meant the agent retained context about me, my work, and my preferences across sessions—not through brittle prompt engineering, but through structured, persistent memory blocks.
Letta's flexibility and SDK customizability aligned well with this project's ethos: user control over everything. It was designed to be extended, not to lock you into a single vendor's workflow.
Thoth is a research assistant built for people who want both power and ease:
It's the tool I needed but couldn't find—so I built it.
# Linux/Mac
curl -fsSL https://raw.githubusercontent.com/acertainKnight/project-thoth/main/install.sh | bash
# Windows (via WSL2)
wsl --install # if needed
curl -fsSL https://raw.githubusercontent.com/acertainKnight/project-thoth/main/install.sh | bash
The installer verifies Docker is running, launches a setup wizard in a container, and installs the thoth CLI command. Takes about 5 minutes.
git clone https://github.com/acertainKnight/project-thoth.git
cd project-thoth
uv sync
export OBSIDIAN_VAULT_PATH="/path/to/your/obsidian/vault"
make dev
thoth start # start services (~4GB RAM)
thoth status # check what's running
thoth stop # shut it down (free RAM)
thoth logs # view logs
thoth update # pull latest version
Thoth is built around loosely coupled services with a shared PostgreSQL+pgvector backend. Letta runs independently from Thoth—restarting Thoth never touches your agent memory or conversation history.
graph TD
User[User]
ObsidianPlugin[Obsidian Plugin]
CLI[CLI]
User --> ObsidianPlugin
User --> CLI
ObsidianPlugin --> APIServer[API Server :8080]
CLI --> MCPServer[MCP Server :8082]
APIServer --> ServiceManager[Service Manager]
MCPServer --> ServiceManager
ServiceManager --> SkillService[Skill Service]
ServiceManager --> LLMService[LLM Router]
ServiceManager --> RAGService[RAG System + Agentic Retrieval]
ServiceManager --> DiscoveryService[Discovery Engine]
ServiceManager --> LettaService[Letta Integration]
SkillService --> BundledSkills[10 Bundled Skills]
SkillService --> VaultSkills[User Skills]
SkillService --> DynamicTools[Dynamic Tool Loading]
DiscoveryService --> Plugins[7 Source Plugins]
DiscoveryService --> AutoScraper[LLM Auto-Scraper]
LettaService --> LettaAPI[Letta REST API :8283]
LettaAPI --> PostgresVector[(PostgreSQL + pgvector)]
RAGService --> PostgresVector
Templates[Template System] --> LLMService
Templates --> RAGService
SettingsJSON[settings.json] -.->|hot-reload| ServiceManager
Agent System — 2 specialized agents (orchestrator + analyst) with 60 MCP tools across 16 categories. 6 memory blocks per agent (persona, preferences, research context, loaded skills, planning, scratchpad). Conversations and knowledge persist across sessions via Letta.
Skill System — 10 bundled skills (paper-discovery, deep-research, knowledge-base-qa, online-research, rag-administration, and more). Users can create custom skills in their vault. Skills hot-reload without restarting services.
Research Discovery — 7 source plugins plus an automated scraper builder that uses Playwright + an LLM to auto-detect article elements on any webpage. Supports scheduled discovery, deduplication, and publication date filtering.
Document Processing — 8-stage pipeline: text extraction, metadata, citations, enrichment, chunking, tags, Obsidian notes, and hybrid index building. Citation resolution runs through a 6-stage chain (Crossref, OpenAlex, ArXiv, fuzzy matcher, validator, decision engine).
RAG & Hybrid Search — Semantic search (pgvector) + BM25 (tsvector) with Reciprocal Rank Fusion. Optional LLM-based or Cohere reranking. Document-aware chunking preserves paper structure. Agentic retrieval mode adds a self-correcting loop with query expansion, document grading, and hallucination checking for complex multi-hop questions.
Configuration — Single settings.json file, hot-reloadable in dev mode (~2 seconds). Everything lives in your Obsidian vault under _thoth/. Change settings through chat or by editing files directly.
Thoth uses Letta for agent memory. You can either use Letta Cloud (hosted, free tier available) or self-host with Docker. See docs/letta-setup.md or docs/letta-cloud-setup.md for details.
project-thoth/
├── src/thoth/ # Python source
│ ├── cli/ # CLI (14 command modules + setup wizard)
│ ├── mcp/ # MCP server + 60 tools
│ ├── services/ # Business logic layer
│ ├── discovery/ # Multi-source discovery + browser auto-scraper
│ ├── server/ # FastAPI REST API (13 routers)
│ ├── analyze/ # Citation resolution + content analysis
│ ├── rag/ # Vector search, reranking, embeddings
│ ├── pipelines/ # Document processing pipelines
│ ├── repositories/ # Data access layer
│ └── .skills/ # 10 bundled skills + role bundles
│
├── obsidian-plugin/ # TypeScript Obsidian plugin
├── templates/ # Jinja2 prompt templates + analysis schemas
├── docker/ # Service-specific Dockerfiles
├── tests/ # Test suite
└── docs/ # Documentation
pytest tests/ # run all tests
pytest --cov=src/thoth tests/ # with coverage
uv run ruff check # lint
uv run ruff format # auto-format
Contributions welcome. See CONTRIBUTING.md for guidelines.
Apache License 2.0 — see LICENSE.
Built with Python 3.12, TypeScript, FastAPI, Letta, PostgreSQL+pgvector, and Obsidian.
An advanced AI-powered research assistant system built for flexibility, extensibility, and user control. Thoth combines intelligent automation with comprehensive customization — every prompt, schema,
總覽
An advanced AI-powered research assistant system built for flexibility, extensibility, and user control. Thoth combines intelligent automation with comprehensive customization — every prompt, schema, and source can be tailored to your workflow.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Project Thoth 透過 stdio 在本機執行。
上次檢查時間:2026-10-26。 An advanced AI-powered research assistant system built for flexibility, extensibility, and user control. Thoth combines intelligent automation with comprehensive customization — every prompt, schema, and source can be tailored to your workflow.