Blog

Connectors and tools: what your model actually sees

A connector is access to a service. A tool is a capability the model can call. The difference shapes what your model sees, and how well it chooses.

2 min read

A grey port block with three slim rods extending from it, one with a green tip

Two words come up constantly when you build with Agent200: connector and tool. They are related, but they are not the same thing, and the difference shapes what your model sees.

A connector is access

A connector is Agent200’s integration with an external service. Depending on the service, it handles:

  • API authentication and the OAuth requirements,
  • access to the service’s supported endpoints,
  • building requests and talking to the provider,
  • retrieving and refreshing tokens,
  • reading or changing data, and returning the results.

The connector is everything you would otherwise write to talk to that provider’s HTTP API yourself.

A tool is a capability the model can call

A tool is what the model sees: a name, a description of what it does and when to use it, and the inputs it accepts. Behind a tool can be one endpoint, several endpoints, or a whole workflow.

Connector:  Slack
Endpoint:   Search messages
Tool:       slack_search

The model never sees the endpoint’s URL, its authentication or its response format. It sees slack_search, decides whether it needs it, and asks for it with the right inputs.

Reading and writing

Connectors are not limited to reading. A connector can expose operations that retrieve messages and operations that send them, depending on what the service’s API offers. Which of those your agent may use is your decision: you enable the endpoints and choose the scopes, and anything you leave off never becomes a tool.

Designing tools the model uses well

Because tools are what the model reasons about, a few habits pay off:

  1. One clear job per tool. “Search messages” is easier to choose correctly than a tool that does everything.
  2. Names that say what happens. A model picks brief or slack_search confidently; it hesitates over vague names.
  3. Only the tools the request needs. Each run() lists its tools. A short list keeps the model focused.
  4. Workflows for fixed sequences. If two calls always happen together, put them in one workflow so the model makes one decision instead of two.

Putting it together

Connectors give your agent reach. Tools give your model choices. Workflows let you shape those choices into capabilities that match your product.

Explore the connectors in the integrations catalog, or read about building workflows.

Stacks of envelopes being organised into an index card cabinet with green tabs
Product

iGPT: your agent’s email context in one request

iGPT is a premium connector on Agent200. It indexes your users' email, threads and attachments, and answers your agent with cited, structured context in…

1 min read

Five frosted glass lenses of different shapes in a row, one with a green edge
Guides

Choosing a web search API for your agent

Tavily, Exa, Brave Search, Perplexity and Firecrawl each lean toward a different part of search. How to pick for your agent, and why you…

2 min read

200 OK

Build the agent.
Access everything it needs.

Bring the agent you already have. Agent200 provides and executes its external capabilities.

Book a demo.
See it in action.

Tell us who you are, then pick a time for a 1:1 session with an expert from our team.

We use your details only to email you about Agent200. See the Privacy Policy.

Pick a time. We'll take it from there.

A 1:1 session with an expert from our team, about what you are building.

Open in Calendly (opens in a new tab)