Built in Tools

English · 简体中文

Built-in Tools

These are the tools the agent has without installing anything. Skills and MCP servers add more on top; this page is the floor.

You never call these yourself — you describe what you want and the agent picks. Knowing the list is still useful, because it tells you what is possible to ask for, and the Provenance panel shows you which ones actually ran.

Files and search

Tool	What it does
read	Read a file from the project. Handles text, and images via view_image.
write	Create or overwrite a file.
edit	Exact string replacement in an existing file.
grep	Regex search across file contents.
search	Find files by name or glob.
view_image	Load an image so the model can actually see it — figures, plots, micrographs.

Running code

Tool	What it does
python	Persistent Python REPL, scoped to the project. State survives across messages.
r	Persistent R REPL. Requires Rscript on PATH and the jsonlite package.
shell	Run a shell command in the project directory.
set_runtime_interpreter	Point a context's Python or R at a specific interpreter, and restart that REPL.

The persistence is the point. python is a kernel, not python -c. Load a 10 GB AnnData object once and keep working with it. Restarting a REPL (via set_runtime_interpreter, or after a crash) clears that REPL's memory — which is also how you recover a wedged runtime without restarting Wisp.

Python provisions a uv virtualenv at .wisp/python/.venv on first use. Wisp never installs R packages automatically.

Remote and long-running work

Tool	What it does
run_in_context	Submit a job to a registered execution context (local, WSL, SSH). Survives disconnects.
get_run	Fetch a run's status and captured output.
monitor_run	Watch a running job. Progress updates do not go through the model.
cancel_run	Stop a run.
transfer_between_contexts	Move files between local and a remote server, or between servers.
configure_ssh_trust	Set up host-key trust for a registered SSH host.

See Execution Contexts for how to register a server and why probing is mandatory.

Web and browser

Tool	What it does
web_scan	Read a page's content.
web_open_tab	Open a URL in a tab.
web_screenshot	Capture a real rendered page as an image the model can see.
web_execute_js	Run JavaScript in the page for inspection.
browser_setup	Prepare the browser bridge.

web_screenshot exists because some pages only make sense visually — a plot on a dashboard, a layout bug, a figure in an online paper.

Delegation and specialists

Tool	What it does
delegate_tasks	Hand sub-tasks to configured sub-agents, in parallel.
get_delegated_result	Collect what a delegated agent returned.
save_specialist	Save the current setup as a reusable specialist.

A specialist is a saved agent configuration — model, tools, instructions — that you can invoke later instead of re-explaining. Delegation runs several of them at once for work that splits cleanly.

Record-keeping

Tool	What it does
research_graph	Record a decision, paper or data asset for the project, or link two of them.
update_plan	Maintain a visible task plan during long work.
attempt_completion	Deliver the final answer for a turn.

What research_graph records shows up in the Research graph panel — see Interface Guide.

Safety boundaries

Worth knowing up front, because they are deliberate and will not be worked around:

• Free-form ssh / scp in shell is disabled. Remote work must go through a registered host, so the alias, user, port and key are the ones you configured. Not the ones a model improvised.
• An SSH host must pass a probe before the agent can use it. Enabling one without a known-good probe opens a dialog telling you to probe first.
• A failed probe, upload or launch is not retried automatically. Wisp opens a connectivity gate for that host and fails further attempts immediately, without contacting the server, until you probe again manually.
• Secrets never touch the project database. API keys and SSH passwords live in the OS keyring.