AI agent tool discovery without loading hundreds of tools
Use progressive disclosure to help AI agents discover, inspect and run the right API tool without loading a growing catalog into context.
By Scrollport
In brief
AI agent tool discovery works better when it starts from the job and loads detail only when needed. Scrollport gives agents nine stable controls: list_apps checks connected-account readiness, search_tools finds a suitable catalog tool, inspect_tool loads its contract and price, run_tool executes it, get_run reads its status and result, get_wallet exposes the human's spending boundary, and get_files, upload_file and delete_file manage workspace references and output files.
AI agent tool discovery should start with the job, not a permanent list of every available integration. Loading hundreds of tool definitions at once consumes context, competes for attention and increases the chance that the model selects the wrong operation. Scrollport instead gives the agent nine stable controls and loads an individual catalog tool only when the job requires it.
AI agent tool discovery starts with the job
A useful discovery layer lets an agent describe an outcome, retrieve a short list of relevant capabilities and inspect the best candidate before execution. A large static tool list reverses that sequence: the agent must compare names, schemas and provider-specific details before it has decided what kind of work it needs to do.
Scrollport keeps the catalog outside the agent’s permanent context. The catalog can grow without requiring users to reinstall a larger bundle of definitions. The agent learns one stable method for finding, understanding and running the right tool.
Nine control tools create a stable interface
The nine control tools are list_apps, search_tools, inspect_tool, run_tool, get_run, get_wallet, get_files, upload_file and delete_file. They are not provider APIs. They are the small control surface an agent uses to reach the wider catalog.
- list_apps checks which human-connected accounts are available to the workspace and ready.
- search_tools searches capabilities and catalog tools using the job the agent needs to complete.
- inspect_tool returns the selected tool’s current instructions, input contract, worked example and price.
- run_tool executes that exact catalog tool with the supplied input.
- get_run reads an existing run, its result, artifacts and cost without starting or paying again.
- get_wallet shows the available balance and the human-controlled spending boundary.
- get_files finds workspace references and generated files, with fresh links and storage usage.
- upload_file adds a reference through an upload link, local client or direct URL.
- delete_file permanently removes an authorised workspace file.
list_apps does not turn account authorization into an agent action. It reports only workspace-visible readiness and, when needed, a Scrollport link for the human to open. The agent never receives the external account identity, connection id or provider credential.
This interface stays the same whether the catalog contains twenty tools or two hundred. Browse the current AI agent tool catalog to see how categories, capabilities, literal tool identities and public providers remain distinct.
Progressive disclosure gives the agent detail when it is useful
Progressive disclosure means revealing information in stages instead of presenting every detail up front. The agent begins with its intended outcome. search_tools returns a short list of relevant choices. inspect_tool then reveals the complete contract for one choice. run_tool is only called after the agent knows the expected input and price.
The sequence reduces context noise without hiding important facts. Literal name, provider, access state, worked examples and billing units still exist; they appear at the point where they can change the decision. The platform documentationexplains the request and response contract for each control tool.
The wallet adds human control to agent execution
Tool selection and spending are related decisions. inspect_tool provides a price estimate before execution. run_tool can proceed beneath the human’s approval threshold or pause for approval when the expected spend is larger. Successful work is settled against one prepaid wallet; failed work releases its hold.
That design lets agents act without granting unlimited authority. The agent receives enough autonomy for routine tasks while the account owner keeps explicit control over larger spends, provider connections and wallet funding.
A growing catalog without a growing setup burden
The practical benefit is not merely a shorter tool list. It is a stable agent connection that can improve as the catalog improves. Newly published tools become discoverable through the same nine controls, while unpublished or unhealthy tools stay out of public discovery.
This is a managed catalog design, not a generic proxy for arbitrary MCP servers. Read the MCP gateway comparison if you are choosing between operating a gateway, connecting servers directly or using a managed catalog. To try Scrollport, follow the quick-start guide or give your agent the setup instructions at scrollport.com/start.