AI Workforce Tools, Connectors, Skills, and MCP

Tools and skills turn AI Workforce agents from chat interfaces into AI automation tools that can do real work. They let agents retrieve trusted information, act in connected apps, follow optimized task playbooks, generate artifacts, and complete longer internal work. Brand Agents use a leaner, customer-facing tool surface. For public agents, see Brand Agent Connectors and Lead Capture.

​

AI Workforce Tool Types

AI Workforce agents can use:

• Knowledge Bases: trusted data from websites, files, text, and Q&A pairs.
• SketricGen connectors: 2,000+ app actions for internal workflows.
• Custom MCP tools: company-specific capabilities exposed through MCP servers.
• API Request: direct HTTP calls for controlled integrations.
• Structured Output: validated JSON for application UI events, typed state, and downstream logic.
• Web Search: current public information.
• Code Interpreter: sandboxed Python for calculations, file processing, and generated artifacts.
• Image Generation: visual asset generation when the tool is enabled.
• Marketplace skills: reusable task packages optimized by other builders.

Use the section links above to jump to the tool type you are configuring.

​

Knowledge Bases

Knowledge Bases let agents retrieve trusted source material from websites, uploaded files, pasted text, and Q&A pairs. In Agent Build, the File Search tool connects one or more approved Knowledge Bases to the agent that needs them. The agent can then retrieve relevant passages during a run instead of relying only on its model knowledge. Use Knowledge Bases when the agent needs approved company, product, policy, support, sales, or website context. For AI Workforce, this usually means internal reference data the run should cite or use while preparing an artifact. For Brand Agents, Knowledge Bases power customer-facing answers from approved website and brand content. Knowledge Bases are retrieval tools, not broad app connectors. If the agent needs to take action in another system, pair the Knowledge Base with a connector, API Request, or custom MCP tool.

​

SketricGen Connectors

SketricGen connectors give AI Workforce agents access to 2,000+ applications. They are useful for internal workflows that need to read or update CRMs, calendars, email, documents, spreadsheets, support systems, project management tools, databases, commerce tools, marketing platforms, and team communication apps. When you add a connector, choose the app first, connect the account, then select only the actions this agent should be able to call. For example, a meeting-prep agent might receive calendar actions to list events and retrieve event details, while a sales operations agent might receive CRM actions to update a record or add a note. Action selection is the permission boundary that matters day to day. Review the selected action names, descriptions, configurable properties, and connected account before running sensitive workflows. For practical workflows, combine connectors with skills. A meeting-prep agent can use a calendar connector to fetch meeting details, a research skill to structure investigation, a document skill to produce the brief, and the run workspace to save the artifact. See Runs, Schedules, Files, and Artifacts for how outputs are stored.

​

Custom MCP Tools

Custom MCP tools let you connect a hosted MCP server directly to an AI Workforce agent. Use this when a service, vendor, or internal platform already exposes an MCP endpoint and you want to make selected MCP tools available to an agent through Agent Build. This is different from a single API Request. A hosted MCP can expose a group of tools, and SketricGen lets you discover those tools, choose the ones the agent can use, and attach the allowlisted set to the right agent. The example below uses a public weather MCP endpoint, fetches the available tools, and allowlists only the weather lookup tool for the agent. To configure a Custom MCP tool:

1. Add a Custom MCP tool node.
2. Enter the hosted MCP server URL.
3. Choose the connection type, usually Streamable HTTP unless the hosted MCP requires SSE over HTTP.
4. Add headers for authentication, such as an authorization token or API key header.
5. Click Filter Tools to fetch the tools exposed by that MCP server.
6. Select only the tools this agent should be allowed to call.
7. Save the tool and connect it to the main agent or the specialist sub agent that needs it.

The allowlist matters. If the hosted MCP exposes ten tools but this workflow only needs two, select those two. That keeps the agent easier to control, reduces accidental tool calls, and makes traces easier to review. Use Custom MCP when you have a reusable hosted tool surface, an internal service that already speaks MCP, or a third-party product that publishes an MCP endpoint. Use API Request when you only need to wrap one REST endpoint. For delegation and routing patterns, see Multi-Agent Orchestration and Handoffs.

​

API Request

API Request turns a cURL-style REST endpoint into an API automation tool the agent can call. Use it when you have one endpoint, webhook, lookup, or action that should become available to the agent without building a full connector or MCP server. The configuration defines the request the agent is allowed to make: HTTP method, URL, headers, query parameters, request body fields, and tool instructions. Those fields become the tool schema the agent follows when it calls the endpoint. The example below wraps a public weather forecast endpoint as a GET tool, then prompts the builder to test the endpoint before saving. You can configure API Request in two ways:

• Manual setup: choose the method, paste the endpoint URL, add headers, define query parameters or body fields, then test the API before saving.
• Import through AI: paste an API documentation URL, OpenAPI spec, or similar reference and describe the endpoint you want. SketricGen maps the endpoint, headers, parameters, and body fields into a usable tool configuration.

In the import flow, paste the API documentation or OpenAPI spec and describe the specific tool you want SketricGen to generate. Use the description field to tell the agent when to call the API and what a successful response means. For example, a lead enrichment endpoint should only be called after the agent has a valid email or domain; a support ticket endpoint should only be called after the user confirms the issue should be escalated. API Request works best for controlled endpoints such as webhooks, internal APIs, simple read-only lookups, enrichment calls, REST API automation, or one-off write actions. If the agent needs many actions from the same service, use a connector or Custom MCP instead.

​

Structured Output

Structured Output lets an agent emit schema-validated JSON when a workflow or application needs predictable data. Add a title, explain exactly when the agent should call it, and define the fields the receiving application expects. Application UI events are one use of the same Structured Output tool, not a separate tool type. For example, a custom frontend can watch for a Structured Output call and use its validated fields to render recommendation cards, open a form, update typed project state, or navigate to the next application step. You can also use Structured Output for classifications, extraction results, connector-ready payloads, and any downstream process that needs predictable fields. The call appears in the run stream so your application can read the arguments and validated result. Structured Output does not automatically pause the run or end the agent’s turn. If the application needs the event to be the final action for that turn, state that explicitly in the agent and tool instructions. Use Human Input instead when the agent must pause and wait for a person’s answer before continuing. For the workflow JSON contract and stream-handling example, see Structured Output for Developers.

​

Web Search

Web Search lets an agent look up current public information when a Knowledge Base is not enough. Use Web Search for market research, recent news, public company information, competitive checks, and time-sensitive facts. For workflows that require approved internal facts, pair Web Search with a Knowledge Base and tell the agent which source should win when sources conflict.

​

Code Interpreter

Code Interpreter gives AI Workforce agents a sandboxed Python environment for calculations, data analysis, file transformation, and artifact generation. Use it when the agent needs to inspect structured data, perform repeatable computation, or create a file rather than only explain an answer in chat. Treat generated files and calculations as run outputs that still require review. Give the agent clear input files, expected output format, and validation criteria, then inspect the resulting artifact and trace before using it downstream.

​

Image Generation

Image Generation lets agents create visual assets when the capability is enabled for your workspace. Use it for workflows that produce social creatives, thumbnails, concept visuals, or supporting images. Treat image-heavy workflows as something to test end to end: the model, tool availability, artifact handling, and any upload connector all affect reliability.

​

Skills Marketplace

Marketplace skills are reusable task packages you can install onto an agent. They make the skills marketplace a practical AI tools marketplace for proven processes such as PDF work, financial analysis, dashboard creation, product requirements, design review, research, writing, or planning. Use skills when you want more than raw app access. A connector lets an agent act in an app. A skill gives the agent a curated way to perform a task well. When browsing skills, review:

• The skill name and description.
• The source or namespace.
• Verification and reputation signals such as installs.
• Category tags that match the work you want the agent to do.
• Whether the skill should be installed on the main agent or a specialist sub agent.

Click Install to attach the skill to the selected agent. After installation, the agent can use that skill during runs alongside its instructions, model, files, tools, and connected apps. Installed skills can be reviewed from the Skills page. This helps teams see which workflows use which skills, check status, edit skill files when needed, and remove skills that should no longer be available.

​

Tool Assignment

Attach tools only where they are needed. The main agent should have the tools required to plan and coordinate the run. Sub agents should receive the smaller tool set required for their specialist task. For example, a research sub-agent may need Web Search and a Knowledge Base, while a reporting sub-agent may need only run files and the dedicated Structured Output tool. For debugging tool calls, action errors, latency, and credit usage, see Traces and Optimization.