Tools, Skills, and MCP

Taskyon treats reusable capabilities as tools. A tool has a name, an LLM-facing description, a JSON Schema parameter contract, and either sandboxed code or a function supplied by its host. It may run locally without an AI, combine deterministic operations with selected model decisions, or orchestrate a fully agentic workflow.

Tools, skills, and protocols

Concept	What it provides	Requires an AI	Requires a server
Skill	Instructions, references, templates, and optional scripts that teach an agent a workflow	Usually, because an agent interprets the instructions	No
Taskyon tool	A typed executable capability with explicit inputs, behavior, and results	No; the implementation chooses whether to call a model	No
MCP	A protocol for discovering and invoking capabilities across a process or network boundary	No	Usually an MCP-speaking process or service

Workflows can also declare temporary scoped tools in their task tree. These definitions apply only to the following branch, do not appear in the saved Tools registry, and stay out of normal model and copied-chat messages. A scoped binding can present a smaller, better-described argument schema to the model while calling an existing registered tool with fixed arguments.

Open Add MCP Tools to fetch or paste an MCP tools/list response, review the mapped Taskyon definition, and save selected tools. A saved MCP definition records the remote server information; it does not make an untrusted server safe.

These concepts can work together. An agent can learn a procedure from a skill and turn its stable steps into a Taskyon tool. Taskyon can import an MCP capability as a tool, and the Taskyon MCP bridge can expose Taskyon tools to another MCP client. MCP defines how a capability crosses a boundary; it does not define whether the capability is deterministic, model-driven, cheap, or safe.

A local tool needs no MCP server or provider request. A tool that calls a remote API still has network cost and exposes its request to that service. An imported MCP tool has the permissions and data boundary of its MCP server.

Model discretion

Tool execution is not limited to an all-or-nothing choice between ordinary code and a fully autonomous agent. A useful model-discretion spectrum is:

• 0: algorithms and commands perform the complete workflow;
• 1–3: deterministic execution handles the normal path, with AI reserved for ambiguity or an optional explanation;
• 4–7: code controls the workflow and evidence, while AI makes selected decisions;
• 8–10: the model chooses most methods and intermediate steps.

This measures model discretion, not operational autonomy. A deterministic tool can run a complete workflow autonomously. Prefer the lowest model discretion that preserves correctness. Moving stable work into code can improve speed, privacy, cost, testing, and replayability.

Record data exposure separately:

• local: data remains in the Taskyon runtime;
• model-visible: selected data is sent to the configured model provider;
• external-service-visible: selected data is sent to an API, MCP server, browser service, or another remote capability.

From conversation to reusable tool

A person and an agent can solve a new problem through research and exploratory tool calls. Once the useful process is understood, they can move repeatable decisions into a typed tool:

1. identify the inputs, outputs, side effects, and blocking ambiguities;
2. research authoritative formats, APIs, or engineering procedures;
3. implement deterministic parsing, validation, commands, and calculations;
4. keep any model call explicit and limited to a documented semantic decision;
5. verify the tool with representative inputs and visible evidence;
6. reuse the installed tool without repeating the original research or model work.

Taskyon agents can perform this process themselves through the tool-creation capabilities described in Agent-Authored Tools. See Tool Examples for ideas across research, daily work, software, science, and engineering.

Tool availability and trust

Open Tools to inspect saved definitions. Open Add MCP Tools to fetch or paste an MCP tools/list response, review the mapped Taskyon definition, and save selected tools. Saving a definition does not make untrusted code, data, or a remote server safe.

Availability depends on the runtime:

• Browser Taskyon provides browser, UI, storage, and sandbox capabilities.
• tycli provides workspace exploration, patching, downloads, documentation search, and shell access inside its environment.
• Embedded applications can register client tools through @taskyon/tyclient.

Developers should use Tool and Workflow Authoring for the current createTool and task-chain contracts.