What Are Agent Tools

Agent Tools are reusable workflows built from ThinkForge skills. You describe a task in natural language, review the proposed steps, and approve the tool before it is saved.

A tool can combine AI processing with deterministic work such as text transformation, data conversion, project-file output, SK Memory search, or connected-service calls. What any particular tool can do depends on the skills in its saved definition and on its configuration.

Packaged Beta Tools

Release builds include a bundle of tool definitions. On startup, ThinkForge adds any missing bundled definitions to the Tool Browser. They appear in the normal browser and folder tree — there is no separate built-in category.

The beta bundle contains seven tools:

  • Create Notion Page
  • Custom Summarize
  • Document Summarizer & Saver
  • Python Script Runner
  • Recent File Monitor
  • Save Doc to Notion
  • Summarize Doc and Save

The bundle installs definitions only. It contains no deployments, triggers, execution history, project state, credentials, or connected-service targets. Python Script Runner starts in the Scripts folder; the other bundled tools start at the browser root. You can organize all of them with the normal folder tree.

Python Script Runner runs one existing local .py file through an explicitly selected interpreter. It uses the normal Agent Tool runtime and does not require an AI model. See Python Script Tools for configuration, context bindings, deployment surfaces, results, and the executable-code trust model.

Provisioning adds only what is missing. A later startup never overwrites an existing tool with the same ID, including its definition, active state, or folder placement. This protects your changes — but it also means an updated bundled definition under an existing ID will not upgrade an installation that already has it.

Tools, Prompts, and Automations

A prompt is an instruction for one AI request.

A tool is a fixed, ordered pipeline of skills. Its steps run in the same order each time. The runtime does not ask an AI model to redesign the plan on every run.

An automation connects multiple existing tools on a canvas. The beta automation runner sorts connected tools and executes that flattened list in sequence. Visual branching, conditions, and merging are not implemented.

What Tools Can Do

Depending on the available skills, a tool can:

  • Process or generate text with an AI model
  • Filter, extract, sort, replace, or reformat text
  • Parse and transform JSON, CSV, and table-like data
  • Search SK Memory
  • Read project input and write output files
  • Call supported connected services after credentials and targets are configured
  • Generate or update supported mind-map content
  • Run a trusted local Python script with context supplied by the calling surface
  • Ask for confirmation or user input when run from an interactive surface

Not every skill in ThinkForge is available to the tool creator. The Capability Manifest is the curated list the planner may use.

Execution Modes

Autonomous runs every step without review pauses.

Guided asks for approval after AI-powered steps, when the calling surface can present one.

A tool's saved mode does not guarantee that every surface can show an interaction. Power Strip, browser-command, and trigger paths run without a person present and automatically approve interaction requests. The Mobile Dashboard has its own approval flow.

Where Tools Currently Run

  • Tools page — create deployments, configure context and model routing, run tools, and manage schedule or file-watch triggers.
  • Power Strip — deploy a tool to an Agent command, then run it with no input or with dropped files, folders, or text.
  • Ribbon — pin a deployment for quick desktop access.
  • Mobile Dashboard — discover active global or current-project deployments, use configured or temporary input, monitor runs, cancel, and answer approvals.
  • Automations tab — use tools as nodes in a sequential automation.
  • Chrome extension — a limited fire-and-forget path can start a tool.

The Chrome extension's Tools tab does not present a browser tool library or return tool results. It opens the desktop Tool Dashboard.

ThinkForge MCP does not expose tool listing or tool execution. It provides project knowledge and document-management tools only.

Creation Flow

  1. Open the Tools page and choose + New Tool.
  2. Describe the workflow in the creation chat.
  3. Answer clarification questions if the planner needs them.
  4. Review the proposed name, description, steps, and skills.
  5. Approve the proposal.
  6. ThinkForge validates and saves the reusable tool.
  7. Use Add from Library to place it in the open project.

The Project | Global switch changes only the deployment inventory. It does not choose where a tool is created. Creation uses AI for analysis and refinement; later runs use the saved linear workflow.

Important Beta Constraints

  • Tool steps do not branch. Chaining and condition fields are stored but ignored during execution.
  • One tool cannot invoke another tool as a step.
  • Saved reference sources are displayed but are not read during execution yet.
  • Missing settings are often discovered only when the affected step runs.
  • A failed step stops the pipeline. There is no built-in retry or resume.
  • A tool can contain only one RunPythonScript step. Use separate script-backed tools in an Automation when a workflow needs several scripts.
  • MCP invocation of user-created tools is designed but not implemented.