74 MCP tools turn a trading question into a research run
Vibe-Trading joins pieces that usually live in separate notebooks and services. A prompt can call market-data loaders, draft a strategy, run a backtest, inspect a document, and save a report with its artifacts. The same workspace is available through a terminal, web interface, REST API, and 74-tool MCP server. That breadth is useful when your problem is the handoff between research steps, rather than one more chart or chat window.
The 68.9 MB checkout contained 2,612 files and roughly 487,439 lines of source at commit 18988fb. That size matches the scope: a Python agent, backtest engines, broker connectors, persistent memory, scheduled research, a React web interface, and several delivery channels all share this repository. You are adopting a platform with many boundaries to inspect, not a thin wrapper around an LLM call.
A 209-package install makes the quick start only the first gate
The README opens with pip install vibe-trading-ai, followed by vibe-trading init. That is a fair path to the CLI, but a useful run still needs an LLM. You can supply a provider key, use ChatGPT OAuth or GitHub Copilot, or point the project at Ollama. Market data has keyless fallbacks, while some feeds and every broker connection bring their own credentials and behavior.
Our environment grew by 209 packages and 800 MB during the 142-second install. Docker packages the backend and frontend together on port 8899, while a source checkout can run the Python API and the frontend separately. The latter requires Node 22.22 or newer. Existing pre-0.1.10 environments may need a rebuilt virtual environment because the project moved to LangChain 1.x.
What happened when we ran it
Our sandbox installed commit 18988fb in 142 seconds and completed its build in 27 seconds. It ran in an unprivileged Debian container with 3 CPUs, 8 GB of RAM, Python 3.12, and no secrets. Pip-audit reported 0 known vulnerabilities. Those results establish that the repository installed and built in the supplied environment; they say nothing about strategy quality or returns.
The test command did not finish inside 900 seconds. Its last output showed passing dots, a few skipped markers, and progress moving from 58% to 62%. The log did not show a failed assertion or an exception before the timeout. We therefore cannot call the suite failed, passed, or broken. The finding is narrower: a complete test result needs more than the 15-minute window we gave it.
Our scan also found 6 CI workflow files, a Dockerfile, and a Compose file, but no top-level tests directory. The long-running command plainly discovered tests elsewhere in the 2,612-file tree. Buyers with a hard release gate should run the exact suite on their own runner and record its completion time before making Vibe-Trading part of a deployment pipeline.
Run cards preserve evidence, but model memory remains an open question
The project is most convincing when it separates an agent's prose from the backtest's own output. Run cards store execution traces and hashes, and v0.1.16 checks cited metrics against the saved CSV before presenting them as grounded. Generated backtest code receives a narrow environment that omits LLM keys, API auth tokens, and broker trading secrets. These controls make a research result easier to audit after the conversation ends.
They do not settle whether an LLM remembered the period it was asked to test. Open issue 1613 proposes recording the model that wrote the strategy and warning when a backtest window ends before that model's training cutoff. Until that work ships, use post-cutoff data, blinded inputs where practical, or independent strategy review. A neat run card cannot remove information that was already inside the model.
Live trading stays behind a broker-specific boundary
Vibe-Trading supports read-only account access, paper profiles, and bounded live order paths for selected brokers. The README says live placement uses broker-specific runtime guards plus limits such as symbol allowlists, exposure caps, and an instant stop. It also states that broker trading is experimental and has not been verified by the maintainers against a real account. That sentence should set your rollout order: research first, paper next, live only after your own checks.
Security defaults are more careful than the phrase "trading agent" may suggest. Non-local web clients require API_AUTH_KEY, and shell-capable tools stay disabled on HTTP and MCP unless an operator enables them. Credentials can live in the OS keyring, while the documentation explicitly warns against pasting broker keys into chat. Those are sensible boundaries, but they still depend on correct local configuration across an 800 MB installed environment.
September 29 activity shows speed and ongoing structural work
GitHub showed 34,341 stars, 15 open issues, and 28 open pull requests when checked on September 30, 2026. The repository was pushed on September 29, the same day v0.1.16 was released. That release reports a large set of data, grounding, channel, and broker changes. The activity is current, and the volume also means buyers should expect frequent behavior changes while the version remains below 1.0.
A maintainer roadmap issue proposes splitting a 3,468-line agent loop that handles tool dispatch, streaming, grounding, compaction, and provenance. That is candid engineering work, and it points to concentration in a heavily changed module. Vibe-Trading earns a trial when one local system connecting research and backtesting solves a real coordination problem. Our 900-second timeout means the trial should include your own full-suite run before any broker account enters the picture.