Write the working rules
Use the instructions builder to define goals, boundaries and deliverables. Use the skill builder for repeatable steps. Add the resulting Markdown to the agent environment you actually use.
Prepare locally. Choose where to run.
Agent Tools: Tokens, MCP, Skills and Offline Evaluation
Use 18 browser-local tools for instructions, skills, context planning, MCP configuration, datasets and deterministic evaluation. Compare official external services by task and requirements.
- browser tools
- 18
- official resources
- 19
Choose the next step for your goal
Align the interfaces
Prepare an MCP configuration and check its structure. Define data shapes with JSON Schema and substitute prompt variables. A valid format still needs compatibility checks in the target client.
Choose a runtime
Choose visual builders for node-based authoring, automation services for business app connections, or code frameworks for state and execution control. The cost tool estimates from prices you enter.
UTools
Local preparation tools
These browser tools generate, check or calculate text and configuration. They need no sign-in and make no model calls. Use them to prepare materials, then choose a runtime to execute agent tasks.
Agent Instructions Builder
Turn your goal, project context and approval rules into an editable AGENTS.md or CLAUDE.md.
Open toolAgent Skill Builder
Create a SKILL.md with safely quoted YAML metadata and your own reusable workflow.
Open toolMCP Config Builder
Build client-specific MCP JSON for Claude Desktop, Cursor or VS Code without running commands.
Open toolMCP Config Validator
Check MCP JSON syntax and supported client semantics, with precise paths for issues.
Open toolPrompt Template Renderer
Fill explicit {{variables}} in prompts, including nested fields and JSON-array batches.
Open toolJSON Schema Builder
Infer a draft 2020-12 JSON Schema from sample JSON and validate another value locally.
Open toolAgent Cost Estimator
Estimate token charges from your own rates, tokens per step, step count and repeated runs.
Open toolAgent Task Planner
Arrange editable task dependencies and inspect a topological order and critical-path estimate.
Open toolToken Counter & Vocabulary Comparison
Count exact o200k_base or cl100k_base text tokens locally, inspect token IDs and compare both vocabularies.
Open toolAgent Context Budget Planner
Allocate a user-defined context window across editable text segments, output reserve and extra overhead.
Open toolRAG Text Splitter with Source Offsets
Prepare traceable text chunks with paragraph, line or fixed boundaries, a hard size cap and adjustable overlap.
Open toolMCP Config Converter
Convert MCP JSON between Claude Desktop, Cursor and VS Code with explicit checks for fields that cannot be preserved.
Open toolMCP Config Merger
Merge 2–10 MCP JSON files with server selection, custom names and explicit conflict strategies.
Open toolMCP Config Diff
Compare MCP servers and fields with stable Added, Removed, Changed and Unchanged results and private values masked by default.
Open toolAgent Output Evaluation Runner
Check existing model outputs with deterministic assertions, Unicode lengths and normalization settings.
Open toolAgent JSONL Dataset Validator
Check text conversations or input/output JSONL with physical line numbers and duplicate-ID detection.
Open toolFunction Calling Tool Schema Builder
Wrap an inline JSON Schema for OpenAI Responses, Chat Completions or Anthropic tools.
Open toolPromptfoo Evaluation Config Builder
Prepare a literal YAML config with prompts, provider IDs, variables and deterministic assertions.
Open toolExternal resource
Official external resources
These links open official websites or documentation in a new tab. UTools does not embed the services or collect their account details or API keys.
Showing 19 of 19 external resources
External resource
Visual builders
Dify
Connect model steps, knowledge retrieval and tools in a visual application workflow.
- Requirements
- Use a Dify Cloud account, or install its server with Docker Compose. Configure the model access and integration credentials your workflow needs.
- Where it runs
- Managed cloud or a server you operate; neither runs inside this directory.
- Data boundary
- Your deployment holds app data. Connected models, plugins and tracing services may receive inputs; self-hosting alone does not keep every request local.
External resource
Visual builders
Coze (global)
Build assistants with prompts, knowledge, tools and workflows on the international Coze platform.
- Requirements
- A Coze account is required. API integrations need access-token authentication and a published agent; service access depends on your account and region.
- Where it runs
- External hosted builder. This entry covers coze.com, separately from the China service.
- Data boundary
- Prompts and uploaded knowledge are processed by Coze and the services you connect. Review those services before adding private data.
External resource
Agent workspaces
扣子 / Coze China
Use a Chinese-language Agent workspace for tasks, projects and collaboration, with skills, MCP and plugins.
- Requirements
- Register or sign in to a Coze China account. The web workspace needs no client installation; individual plugins can require extra connections or permissions.
- Where it runs
- Hosted workspace with separate desktop and mobile clients; not a browser-local utility.
- Data boundary
- Projects, conversations and files sync through the account. Connected plugins and MCP services add their own data access boundaries.
External resource
Workflow automation
n8n
Connect triggers, business apps and AI steps into repeatable workflows.
- Requirements
- Use an n8n Cloud account or install and maintain a server. App connections require their own credentials; AI model nodes need suitable model access.
- Where it runs
- Managed cloud or self-hosted service. License and edition terms apply to your use case.
- Data boundary
- Workflows can move data between connected apps. Hosting controls where n8n runs; each destination still receives the data sent to it.
External resource
Visual builders
Flowise
Prototype agents and LLM flows with a visual node editor and reusable components.
- Requirements
- Use Flowise Cloud or install the Node.js/Docker server. Configure credentials for the model, vector store and tools you choose.
- Where it runs
- Cloud or a maintained backend; the editor is not an offline-only web tool.
- Data boundary
- The Flowise server and configured services process your flow data. Check storage, model and tool destinations before using confidential content.
External resource
Visual builders
Langflow
Connect visual components into agent flows, then expose a flow through an API.
- Requirements
- Install and start Langflow, configure server sign-in and your model credentials. API execution also uses a Langflow API key; the official tutorial uses OpenAI.
- Where it runs
- Run the Python service locally or deploy it on your infrastructure.
- Data boundary
- Your Langflow server stores flows; connected components can send data to model and third-party APIs. Server installation does not imply offline inference.
External resource
Code frameworks
OpenAI Agents SDK
Write agents, function tools and handoffs in JavaScript or Python while controlling the application runtime.
- Requirements
- Install the SDK in a development project. The official quickstart requires an OpenAI API key and API access; your application provides deployment and storage.
- Where it runs
- Application library, not a hosted visual builder. Install and run it in your own runtime.
- Data boundary
- The SDK runs in your app, while model requests go to the configured provider. The standard server-side path includes tracing; review tracing and data settings.
External resource
Code frameworks
Google Agent Development Kit
Build tool-using agents and agent teams in code, and test them with a local CLI or development web UI.
- Requirements
- Install ADK in a supported language runtime. The Python quickstart uses a Gemini API key; other providers and cloud deployments have their own credentials.
- Where it runs
- Developer framework with local testing and separate deployment options.
- Data boundary
- The local development UI runs an agent runtime, not a local model by default. Data can go to the model, tools and deployment services you configure.
External resource
Code frameworks
CrewAI
Coordinate role-based agents and use Flows to control state and execution order.
- Requirements
- Install a Python environment and CrewAI CLI. Configure your LLM credentials; the official research tutorial also needs a Serper search API key.
- Where it runs
- This entry covers the installable framework; managed CrewAI offerings are separate.
- Data boundary
- Agents can send content to configured models, search and tool APIs. Review optional tracing and the services used by each task.
External resource
Code frameworks
LangGraph
Control stateful agent graphs with durable execution, streaming and human review points.
- Requirements
- Install the framework and write your graph. A mock graph can run without a model key; real model calls need configured provider access. LangSmith is optional.
- Where it runs
- Installable orchestration runtime; hosted deployment and observability are separate choices.
- Data boundary
- You choose model, tool, checkpoint and tracing destinations. Enabling a hosted tracing service sends run details beyond your runtime.
External resource
Coding agents
OpenAI Codex
Explore a codebase, make changes and run development tasks through local clients or cloud workspaces.
- Requirements
- Local clients require installation and supported sign-in: ChatGPT or an API key. Codex cloud requires ChatGPT sign-in; plan and workspace permissions govern access.
- Where it runs
- Local app/CLI/editor clients and cloud workspaces; local execution still uses configured model services.
- Data boundary
- Local files can become model context. Data settings depend on whether you use ChatGPT or API authentication, your workspace and connected services.
External resource
Coding agents
Claude Code
Work on a code project through a terminal agent that can inspect files, edit and run commands.
- Requirements
- Install the CLI and sign in using a supported Claude subscription, Claude Console account or supported cloud provider. You also need a code project and terminal.
- Where it runs
- This entry covers the local CLI described in the official quickstart.
- Data boundary
- Project context is sent to the configured model service. Review repository permissions, tool approvals and the data rules for your chosen account or provider.
External resource
Testing and evaluation
Promptfoo
Compare prompts and model outputs with datasets and configurable assertions.
- Requirements
- Install the CLI to run evaluations. Model-backed tests require provider access; execution can incur provider charges. Our configuration builder does not execute it.
- Where it runs
- Local CLI and optional sharing services; not embedded in UTools.
- Data boundary
- Promptfoo runs separately; selected model providers can receive prompts and test cases. Review telemetry and provider settings before use.
External resource
Tracing and observation
Langfuse
Trace agent runs, manage prompts and organize evaluations across your application.
- Requirements
- Use a cloud project or deploy the server, then connect your app through the SDK or supported integrations. Access and hosting costs depend on your setup.
- Where it runs
- Hosted project or a maintained self-hosted service.
- Data boundary
- Configured tracing can transmit prompts, outputs and metadata to your Langfuse deployment. Set redaction and access rules before collecting private data.
External resource
Tracing and observation
LangSmith
Inspect production traces and build datasets for debugging and evaluation.
- Requirements
- The documented hosted quickstart requires an account, project and API key, plus an SDK or framework integration. Check your plan and limits directly.
- Where it runs
- External hosted tracing; no connection has been configured here.
- Data boundary
- Tracing sends configured application activity to the service; decide what may be captured and who may view it.
External resource
Testing and evaluation
Ragas
Evaluate RAG and agent workflows using task datasets and explicit metrics.
- Requirements
- Install the Python library and prepare suitable datasets. Model-graded or embedding metrics need the corresponding provider or local model runtime.
- Where it runs
- A development library, not a browser-only evaluator or a connected integration.
- Data boundary
- Data handling depends on the chosen metrics and model providers. A local Python process may still call external models.
External resource
Knowledge and retrieval
LlamaIndex
Build ingestion, retrieval and agent workflows with explicit data connectors and storage.
- Requirements
- A development environment is required. Choose and configure the data stores, embedding models and language models your workflow uses.
- Where it runs
- A framework for an application you run; our text splitter only prepares local chunks.
- Data boundary
- Connectors, storage, embeddings and language models each have their own data boundary. Review the complete path before adding private documents.
External resource
Browser and web data
Firecrawl
Convert selected web pages into structured or model-readable content through a separate service.
- Requirements
- The hosted API needs an account and API key. Check endpoint support, credits and target-site permissions before making requests.
- Where it runs
- External web-data service; its API is not connected to UTools.
- Data boundary
- Requested URLs and extracted content go through the chosen service. This directory never submits your URLs to Firecrawl.
External resource
Browser and web data
Browser Use
Run browser tasks through a hosted agent, cloud browser or separately installed library.
- Requirements
- Cloud use needs an API key; a local setup needs the library, browser and chosen model access. Check the official setup and billing for the selected path.
- Where it runs
- Cloud paths and a separate open-source library; no browser automation is launched here.
- Data boundary
- Browser sessions may expose page content and account state to the chosen runtime and models. Scope tasks and permissions in that platform.
Check three things before choosing a platform
Who operates the runtime
Hosted services start with an account. Self-hosting requires maintaining a server and storage. Code frameworks need a development environment. A local installation does not imply local model inference.
Where the data goes
Check model, plugin, search, MCP and tracing services. Self-hosting controls where the application runs; connected external services may still receive your inputs.
How you will verify it
Start with one small task to check tool calls, failure handling and human review points. Test exported configuration in the target platform and compare cost estimates with actual billing.
Scope: public official documentation and destinations checked on 2026-11-03. We did not sign in, install or run these services. Plans, regions, permissions and features can change; check the official pages. Resources are grouped editorially, without ratings or market rankings.
OpenAI Agent Builder is in a deprecation transition, with a scheduled shutdown on 2026-11-30 in the official documentation. This directory lists Agents SDK for developers who want to control agent workflows in their application. Read the official deprecation notice