Tools

Shinbo's 27 tools and the seven that ask

Shinbo adds 27 tools of its own to the harness, seven of which ask before they act, and can write new ones mid-conversation.

These guides describe the current development source. The published v0.7.1 release can differ; check release notes before following a workflow.

Twenty-seven of Shinbo's own. The shell is the harness's.

Shinbo ships no bash of its own. Files, search, shell, language-server queries, skills, MCP and subagents belong to the harness, shinbo-cli, and are gated over one permission channel. Shinbo's 27 tools are appended to that same registry, so the model sees one list and one set of rules.

Try it

Open Settings → Tools. Built-in tools are grouped by what they do, each group with an enabled count.
Expand a group and switch off any tool you do not want offered. The hide applies in every permission mode, and a model that guesses the name is refused.
Scroll down to Added to Shinbo: Shinbo's own written tools, imported skills and MCP servers, each with its own count.
Start a thread and ask Shinbo to write a tool or skill for something you do often. It is live on the next turn.

Settings, Tools: new-thread mode above grouped built-in tools with enabled counts, followed by installed tools, skills and MCP server groups — **Tool settings** — Expand a group to choose tools · installed additions have their own countsCaptured September 12, 2026 · current app source · example profile

The 27, and which ones ask

Seven tools carry an ask gate. In Ask and Accept edits mode they stop for you every time. In Auto the verifier model clears ordinary calls and hands you anything it will not clear. In Full access they run on their own, except computer, which always asks per app. Everything marked auto runs without a prompt in every mode. Modes are described under Control.

browserask

cliask

cli_runsauto

computerask

shortcutauto

write_skillauto

write_toolauto

write_pluginauto

run_toolask

memoryauto

advisorauto

visionauto

secretask

web_searchauto

task_listauto

planauto

goalauto

threadsauto

read_traceauto

contextauto

keepauto

agentsauto

install_mcpask

workflowask

artifactauto

componentauto

visualizeauto

Shinbo writes its own

A handful of those tools extend the registry itself. Anything they install mid-turn is available on the next turn, with no restart.

write_skill: records a lesson as a skill the model can load later
write_tool: writes a script and registers it as a tool callable by name
run_tool: runs one of those scripts — ask-gated
install_mcp: adds an MCP server — ask-gated
write_plugin: packages skills in the Codex plugin format
shortcut: binds a prompt to a key combination; three slots

Web search

web_search can go through TinyFish, 4get, SearXNG, Brave Search, Tavily or Exa. The provider that answered shows its logo in the transcript. Search goes to that provider even when the thread's model is local.

Inspecting tool calls

Every tool call is a span in the thread's trace. Read it in the inspector, or let the agent read it with read_trace, which takes a numeric offset and returns up to eight traces per call. Recorded thread data exports as CSVs. Retention is covered under Self-improvement.

The same agent, headlessgated on the tty

$ shinbo-cli ask "explain this repository" one-shot answer $ shinbo-cli a REPL in the current directory $ shinbo-cli session resume last

Outside the app the harness asks on the terminal instead of over Shinbo's permission channel. The v0.7.1 release still ships the binary under its old name, emma-cli, inside the app bundle.

Limits

A hidden tool stays hidden in every mode, including Full access; hiding is not a per-mode setting.
computer asks per app in every mode and runs under fixed rails; see Control.
An installed MCP server or written tool is only as trustworthy as its source; install_mcp and run_tool ask for that reason.

Related: Control for the modes behind each gate, Harness for CLI adapters and imports, Notch & surfaces for artifacts that replace interface regions.