Your Investment Research Stack, Running on Your Own Machine

You've probably used a cloud-based AI assistant for something work-related and had that nagging thought: where does all this data actually go? For most tasks, you shrug it off. But when the data is your portfolio positions, your investment thesis, and your research conversations, the question gets harder to ignore. HunterCode is an open-source, self-deployed AI research assistant for A-shares, Hong Kong stocks, and US equities that keeps your data, conversations, and positions on your own machine.

What It Does

HunterCode is a financial AI assistant that runs on your own computer or server. You give it a large language model key, and it handles market data queries, news retrieval, individual stock deep-dives, trend prediction, and watchlist/portfolio management. It works according to SKILL files you write yourself, which means you define the analysis methodology rather than accepting whatever the tool decides.

Architecturally, it's a Docker-based deployment with six services, including PostgreSQL and Redis. Images are available for both amd64 and arm64, so Apple Silicon and ARM cloud hosts run natively without emulation. You can deploy it locally with Docker Compose, or push it to a cloud platform like Zeabur, Sealos, Railway, 1Panel, Coolify, or Dokploy if you don't want to manage a server. Conversations, positions, and investment theses are stored in a local database.

The project positions itself as an open-source local alternative to Tencent's WorkBuddy financial edition, aimed at private funds and professional individual investors. It doesn't connect to brokerages for trading, doesn't make decisions for you, and doesn't guarantee prediction accuracy. It's a research tool that organizes public data, analysis methodology, and LLMs into one place.

Why It's Cool

• Data stays local by default. Your conversations, holdings, and investment arguments live in a local database. If you're managing other people's money or just prefer not to hand your positions to a cloud vendor, this matters.

• Built-in LLM quota means zero setup friction. You can apply for a free hunt_tools_ platform key (about 30 seconds), and the setup wizard auto-fills the endpoint and model name. That's 300,000 tokens per day, resetting at midnight Beijing time. If you'd rather use your own key, DeepSeek or any OpenAI-compatible gateway works too, and you can switch between the two at any time.

• The privacy claim is specific. The gateway only logs token counts and model names, not prompts or responses. That's a concrete commitment, not a vague "we care about privacy" statement.

• SKILL files are the interesting part. Instead of a fixed analysis pipeline, you write your own methodology and the assistant follows it. For anyone with a repeatable research process, this is the difference between a toy and a tool.

• Cloud deployment templates come with honest caveats. The README states plainly that there's no one-click deploy button because the maintainers don't have accounts on those platforms and haven't done a real deployment. Instead, they provide templates validated by mechanical translation to Compose, running the full flow locally from empty volumes. Each doc says what was tested, what wasn't, and what you need to verify yourself. That level of transparency is rare.

• Public deployment is handled thoughtfully. Instances on cloud platforms are public by default, so templates disable single-user no-login mode and generate an initialization token. Without it, whoever opens the instance first could configure their own LLM. The token is visible in the platform's environment variable panel.

How to Try It

You'll need Docker Desktop (Windows/macOS) or Docker Engine with Compose v2 (Linux), 10 GB of disk, 4 GB of RAM (measured peak is about 1.3 GB), and access to ghcr.io.

1. Get an LLM key. Either apply for the free built-in quota key at hunter.agentpit.io/dev/api-keys, or grab a key from DeepSeek or any OpenAI-compatible gateway.

2. Clone the repository and start the stack.

git clone https://github.com/agentpit-io/hunter-community
cd hunter-community
docker compose up -d

1. Run the setup wizard. Open the local URL and walk through the steps. If you're using the built-in quota, select the first card in step 2 and the endpoint and model name fill in automatically.

2. Write a SKILL file if you want the assistant to follow your own analysis methodology.

3. Explore the online demo at hunter-community.agentpit.io if you want to see it in action before committing to a local install.

Full documentation is in the repo's docs/ directory, and there's a dedicated getting-started guide. If you deploy to a cloud platform and it works, the maintainers explicitly ask you to report back in the Issues section, since they haven't been able to test those paths themselves.

Final Thoughts

HunterCode is aimed at a specific audience: individual investors and small fund researchers who are comfortable with Docker and want control over their data. If you fall outside that group (say, you want a one-click web app with no setup), this isn't for you yet. But the combination of local data storage, custom SKILL files, and a built-in LLM quota removes two of the biggest barriers to self-hosted AI tooling: cost and configuration. The maintainers' willingness to document what they haven't tested is a good sign for a project that's still early. Worth watching, and worth trying if you've been looking for a research assistant that doesn't require trusting someone else's server.

Follow @githubprojects for more developer tools and open source projects.