Self-hosted financial AI that remembers why you bought

What it does

Hunter is a self-hosted financial AI platform that wraps a multi-agent architecture around standard LLM APIs. You provide one model key and it orchestrates real-time quotes, news, technical analysis, and deep research workflows through 23 built-in financial SKILLs, streaming results back in a web UI. Everything from conversation history to portfolio positions and investment theses lives in a local Postgres database, so your data never leaves the machine.

The interesting bit

The standout feature is “Investment Thesis” memory: the system records your entry rationale, cost basis, and key assumptions, then continuously monitors incoming data to flag when pillars of your thesis start to crack. It is essentially an automated “why did I buy this?” audit that outlives the initial report.

Key highlights

• Runs entirely via Docker Compose and serves a web UI locally; the README claims a five-minute startup once images are cached.
• Three-tier data supply: free built-in sources (akshare, yfinance), bring-your-own MCP tools or broker APIs, or a platform-issued key that unlocks 32 data sources plus Kronos time-series prediction and TrueSource alternative intelligence.
• Pluggable LLM backend with tested compatibility matrices; DeepSeek v4 pro is the default P0 choice, though Claude Sonnet 5 and Qwen 3.8 Max score higher on tool-calling reliability.
• Built-in SKILLs are written in Markdown and cover due diligence, valuation, and portfolio methodology; new ones can be installed from GitHub without restarting the core.
• Apache 2.0 license; the authors explicitly permit commercial forks as long as you rebrand.

Caveats

• Model tool-calling reliability is uneven: DeepSeek v4 pro hits only 6/7 golden cases, GPT-5.6 sol hits 5/7, and Gemini 3.5 Flash can timeout on boundary calls. DeepSeek also requires LLM_SCHEMA_SANITIZE=1 to behave correctly.
• The free data tier has incomplete coverage and will explicitly downgrade when sources are missing, so serious use likely requires applying for the platform pipeline key or wiring your own MCP servers.
• Resource appetite is nontrivial: the opencode image alone is roughly 7.5 GB, and the authors recommend 20 GB of disk and 4 GB of RAM.

Verdict

Worth a look if you are an individual investor who wants AI research assistance but refuses to let chat history or portfolio logic touch a SaaS database. Skip it if you need battle-tested execution infrastructure or are unwilling to babysit LLM tool-calling quirks and data source gaps.