ValueCell: a multi-agent financial platform you run on your own machine

The gap ValueCell is trying to fill

Most retail trading tools hide their reasoning. You get a signal, a chart overlay or an alert, and no way to see which input produced it. ValueCell takes the opposite position: it exposes a team of agents, each with a stated job, and lets you watch them work. The README describes a DeepResearch Agent that retrieves and analyzes fundamental documents, a Strategy Agent that executes multi-strategy crypto trading, and a News Retrieval Agent that pushes scheduled news. The stated mission is to build a decentralized financial agent community, and the licence is Apache-2.0, so the code is inspectable rather than a black box.

The audience is narrower than the marketing suggests. You need Python 3.12 or newer to run from source, an API key from at least one supported model provider, and, if you want live execution, exchange credentials. The README also notes that the hosted product at valuecell.ai covers A-share deep research and market analysis without deployment, which means the self-hosted path is for people who specifically want the agents running on their own hardware. The README claims sensitive information stays local, and the Live Trading notice repeats that the app stores secrets locally. That is a design promise, not a security audit, and the repository ships a SECURITY.md for anyone who wants to check further.

How the agents, models and exchanges connect

The architecture image in the README sits under Key Features, and the surrounding text gives the pieces. On the model side there are multiple providers: OpenRouter, SiliconFlow, Azure, an OpenAI-compatible endpoint, Google, OpenAI and DeepSeek. On the data side the README lists US, crypto, Hong Kong and China markets. On the framework side it names Langchain and Agno, connected through the A2A Protocol, which is what makes the agent layer pluggable rather than monolithic.

The exchange layer is where the design gets concrete. Binance uses USDT-M futures and the international site only, not the US site. Hyperliquid uses USDC as margin currency and authenticates with a main wallet address plus an API wallet private key; market orders are converted to IoC limit orders, and the minimum is 10U per trade. OKX requires an API key, secret and passphrase. Trading pair formats differ by venue: BTC/USDT on Binance and OKX, but SYMBOL/USDC on Hyperliquid, which the README says must be adjusted manually.

The .env.example file shows the configuration surface. Model keys live there, and a comment states that to access full functionality you should include an embedding-capable provider, with the note that OpenRouter does not currently support embedding or reranker models. That is a real constraint on the default setup: the primary provider cannot cover the embedding path by itself.

Installing ValueCell and configuring a first model

The README splits the path by user type. New users are told to download the latest MacOS or Windows application from the GitHub Releases page, install it, then configure a model provider before first use. If you would rather run from source, the repository is a Python project with a python/ directory, a frontend/ directory and a Makefile at the root. The Makefile targets assume uv and ruff are available.

make format
make lint
make test

Those three targets run ruff format and isort over ./python, ruff check over the same tree, and pytest through uv run. They are the contributor loop, not the runtime. For runtime configuration, copy the example environment file and fill in keys.

cp .env.example .env

Inside .env, the API block sets the host and port the backend listens on, and the model block takes provider keys. The example values are API_HOST=localhost and API_PORT=8000.

OPENROUTER_API_KEY=
OPENAI_API_KEY=
SILICONFLOW_API_KEY=

The file also carries APP_ENVIRONMENT=development, API_DEBUG=true and AGENT_DEBUG_MODE=false. Leave AGENT_DEBUG_MODE off unless you are diagnosing an agent; the README does not describe what verbose agent output looks like, so treat it as an unverified switch. After the model key is in place, the README says to add the AI model API key through the web interface, which is the same step the desktop build asks for on first launch.

Live trading is contract-only and unevenly verified

The README is unusually direct about this. The notice states that ValueCell currently supports leverage trading only, that spot is implemented as 1X contracts, and that you must ensure your Perps account has sufficient balance. A separate note warns that API secrets must be kept secure to avoid losing funds. If you want plain spot exposure without a derivatives account, this is the wrong tool, and no configuration flag changes that.

The exchange table is the second constraint. Binance, Hyperliquid and OKX are marked Tested. Coinbase, Gate.io, MEXC and Blockchain are marked Partially Tested, which the legend defines as code implementation complete but not fully tested, possibly requiring debugging. The README recommends prioritizing the fully tested venues. Treat the Partially Tested rows as code you may have to fix, not as supported integrations.

Each venue carries its own sharp edge. Binance requires the international site and a non-zero perpetual contract balance; the README suggests adding an IP whitelist when applying for the API key. Hyperliquid needs USDC margin and a manually adjusted pair format. OKX needs the account passphrase as a third credential. None of these are documented as interchangeable, and the README does not describe a rollback or kill-switch procedure for a running strategy beyond the start/stop control in the interface.

Where ValueCell sits next to TradingAgents

TradingAgents appears in the search data alongside ValueCell, and the two overlap in framing but not in scope. TradingAgents is a research-oriented multi-agent simulation: agents debate a thesis and produce an analysis. ValueCell keeps that research layer, in the DeepResearch Agent, and adds an execution layer that routes live orders to OKX and Binance with what the README calls built-in guardrails.

That difference decides your choice. If you only want to study how an LLM pipeline reasons about a ticker, a research-only framework is the smaller dependency and carries no exchange risk. If you want the same reasoning to end in an order on a derivatives account, ValueCell is the one that closes the loop, and it also takes on the operational burden: credential storage, pair formatting, margin currency and the contract-only restriction. The trade is explicit. You get a shorter path from analysis to execution, and you accept a larger blast radius when a strategy misbehaves.

Maintenance, upgrade cost and the Apache-2.0 licence

The last push to the default branch was on 2026-03-09, and the most recent release listed is v0.1.20, tagged ValueCell-0.1.20-beta on 2026-01-10. Every release in the list carries a beta suffix, which is the project telling you what stage it considers itself to be in. The repository is not archived, but with the last push more than six months before today, calling it actively developed would overstate the evidence. Plan for the code you clone to be the code you maintain.

Upgrade cost is dominated by the provider matrix. The .env.example comment about SiliconFlow is a concrete example: if you obtained your key from siliconflow.com rather than siliconflow.cn, the README says you need to update siliconflow.yaml manually. Model provider defaults change faster than the application does, so budget time for configuration drift on each upgrade rather than expecting a clean pull.

The licence is Apache-2.0, which permits commercial use and modification with the usual attribution and notice requirements. It grants no warranty, which matters here because the README itself states that investing involves risk. That is a factual statement about the software, not legal advice, and anyone deploying this against a funded account should read the licence and SECURITY.md directly rather than relying on a summary.

Editorial conclusion

ValueCell fits developers and technically comfortable traders who want to inspect how an agent pipeline reaches a trading decision, and who are willing to run it locally with their own model and exchange keys. It does not fit anyone who wants a managed brokerage product, anyone trading spot only, or anyone unwilling to read the exchange notes before funding an account. Before committing money, verify three things: that your exchange appears as Tested rather than Partially Tested in the README table, that your model provider supports embeddings, and that you understand the contract-only execution model described in the Live Trading notice.

Frequently asked questions

Official sources

1. License: Apache-2.0
2. Project website
3. README
4. Releases
5. ValueCell-ai/valuecell on GitHub