Tools
Every agent has the same base kit, five finance tools, and BytNode. Exchange tools come with the accounts you attach.
The base kit
Read | Reads text, images, PDFs and notebooks. |
Write, Edit, NotebookEdit | Creates and changes files. Edit replaces one exact string. |
Glob, Grep | Finds files by name, searches their content (ripgrep, embedded). |
Bash, PowerShell | Runs a command in the working folder, bounded in time. |
Monitor | Watches a command or a WebSocket for a while, and reports. |
WebSearch, WebFetch | Searches the web; fetches a page as Markdown. |
Workflow | Delegates tasks to sub-agents, each with a profession. |
Skill | Opens one of your skills and follows it. |
Finance
SaveMarketData | One BytNode tool call written straight to a CSV in the working folder, without pushing the rows through the model. |
Chart | A PNG from a CSV: candles or line, volume, overlaid columns, markers, levels. The agent sees the image; so do you, in the thread. |
Backtest | One standard engine in Python: next-candle execution, fees, slippage, funding, position sizing by risk, in and out of sample, metrics per period. The agent only writes signal(df) and stop_distance(df). |
Memory | Dated notes in the working folder; the last 25 are slipped into every prompt. |
TodoWrite | The agent’s own task list on a long request, shown in the thread. |
Sub-agents
Workflow splits a long request into tasks. Each task names a profession and a model; sub-agents run three at a time, eight at most, with the profession’s skills and every tool but Workflow and the exchange actions. They ask no permission, and each returns a report signed by its profession, merged into the answer.
Permissions
In Ask before acting mode, Bash, PowerShell, Monitor, Write, Edit, NotebookEdit, Backtest and Workflow wait for your answer in the thread: Autoriser, Toujours pour cet agent (a rule for that tool or that command prefix), or Refuser. An irreversible command, a recursive delete or a force push for instance, never gets a Toujours. Exchange orders, closes and transfers always ask, in every mode.
Context, handled for you
Every tool call is checked against its schema before it runs; a malformed call goes back to the model with the reason. Past 70 % of the model’s window, old tool results are pruned from what the model receives (the thread you see keeps everything); at 90 % the thread is compacted; if the provider refuses the context, it is compacted and your message replayed.