No-code plugins

Wire up tool calls with a few lines of YAML

Give an assistant your database, your web services, the open web or your own procedures. Describe the tool in a manifest, apply it, and list it in an assistant. There are no functions to write, no JSON Schemas to maintain and nothing to deploy.

How it works

From a manifest to a tool call, in three steps

A tool call is how an AI model asks your systems for something it doesn't know. Building one by hand means a function definition, the code behind it, its credentials and its error handling, for every tool. A plugin is a description of the tool. Smarter does the rest.

  1. 1

    Describe the tool

    Say what it is for, name its parameters in plain words, and point it at the data. That is the whole job.

    stackademy-sql.yaml
    kind: SqlPlugin
    metadata:
    name: stackademy_sql
    spec:
    connection: stackademy_catalog_db
    sqlData:
    sqlQuery: SELECT ... {max_cost}
  2. 2

    Apply it

    From the command line, the API, the web console, or a pull request to a repository of manifests.

    terminal
    $ smarter apply -f stackademy-sql.yaml
  3. 3

    List it in an assistant

    Add its name to an LLMClient. The model is offered the tool, and Smarter runs the calls it makes.

    stackademy-advisor.yaml
    kind: LLMClient
    spec:
    plugins:
    - stackademy_sql
New to tool calling? Seven minutes on what it is, and why a question about data that changes, like the weather forecast, needs a tool, with Smarter showing each tool call as it happens.

Off the shelf

Six plugins, ready to configure

Each is a kind of manifest. Pick the one that matches where the answer lives, fill in the details, and your assistant has a new tool.

kind: SqlPlugin

Answers from your database

Write one parameterized query. The model never sees it: it sees the parameters, and fills them in from the question. Smarter runs the query over a Connection to MySQL, PostgreSQL, SQL Server, Oracle and others, and hands back the rows.

Ask it“My budget is $1,000. Which AI courses can I take?”
  • Every value is inserted as a quoted, escaped literal, so a question can never become SQL.
  • An enum limits a parameter to the values your data actually holds.
  • The database password is a Secret, never part of the manifest.
  • A row limit, and test values to try the query before any model calls it.
SqlPlugin documentation
stackademy-sql.yaml
connection: stackademy_catalog_db
sqlData:
sqlQuery: SELECT c.course_code, c.course_name, c.cost
FROM courses c
WHERE (c.description LIKE CONCAT('%', {description}, '%')
OR {description} IS NULL)
AND (c.cost <= {max_cost} OR {max_cost} IS NULL)
parameters:
- name: max_cost
type: number
description: the maximum cost that a student is willing to pay.
- name: description
type: string
enum: [AI, mobile, web, database, network]
limit: 10
The SqlPlugin in 60 seconds: an assistant answers from a database, through a query declared in a manifest.
kind: ApiPlugin

Calls your web services

Point the model at an endpoint of a system you already run: an order lookup, an inventory count, a course catalog. The values it chooses become the request’s query string or body, and the response becomes its answer.

Ask it“Which web development courses do you offer?”
  • Typed, described parameters, with enums, required fields and defaults.
  • The base url and its credentials live in a Connection and a Secret.
  • Any method, headers, url parameters and a JSON body.
  • The same parameters as a SqlPlugin, so a tool can move from a database to an API.
ApiPlugin documentation
stackademy-api.yaml
connection: stackademy_catalog_api
apiData:
endpoint: /api/v1/demos/stackademy/course-catalogue/
method: GET
parameters:
- name: max_cost
type: number
description: the maximum cost that a student is willing to pay.
- name: description
type: string
enum: [AI, mobile, web, database, network]
limit: 10
kind: Plugin

Answers from facts you write down

Office hours, a price list, contact details, this month’s promotions: write them in the manifest as plain YAML. The assistant answers from them, exactly, instead of from whatever the model remembers.

Ask it“Do you have any coupon codes?”
  • No database, no connection, no credentials. Just the facts.
  • Changing an answer is editing a line of YAML.
  • Offered only when the question mentions it, with search terms.
Plugin documentation
everlasting-gobstopper.yaml
selector:
directive: search_terms
searchTerms: [Gobstopper, Gobstoppers]
data:
staticData:
contact:
- name: Willy Wonka
- phone: +1 123-456-7890
couponCodes:
- code: 10OFF
description: 10% off your next purchase
kind: SkillPlugin

Follows your procedures

Give the assistant an Agent Skill: a SKILL.md of instructions, with templates and examples beside it. Write it in the manifest, or name a skill in a GitHub repository, such as Anthropic’s public skills, and Smarter reads it from there.

Ask it“Here’s the transcript of today’s standup. What are the action items?”
  • The open SKILL.md format, so a skill written for one tool works here.
  • Resources, such as a template, travel with the skill.
  • Or a source url: the skill stays in its own repository.
SkillPlugin documentation
skill-meeting-notes.yaml
skillData:
skill: |
---
name: meeting-notes
description: Turns raw meeting transcripts into structured notes.
---
1. Read the entire transcript before writing anything.
2. Fill in [the notes template](assets/template.md).
3. Every action item must have an owner. Never invent one.
resources:
assets/template.md: |
## Summary ## Decisions ## Action items
kind: WebsearchPluginExperimental

Researches the web, and cites it

Let the assistant search the web, read the pages it finds, and answer with the sources it used. You choose the search provider, how many results, which sites are off limits, and how long results are cached.

Ask it“What changed in the latest release of PostgreSQL?”
  • Allowed and blocked domains, and safe search.
  • robots.txt is obeyed, and pages are trimmed to a size you set.
  • The search api key is a Secret.
WebsearchPlugin documentation
websearch-research-assistant.yaml
websearchData:
search:
provider: brave
apiKey: brave_search_api_key
maxResults: 5
safeSearch: moderate
fetch:
maxCharacters: 20000
respectRobotsTxt: true
blockedDomains: [pinterest.com, quora.com]
cacheTtl: 3600

Included with every plugin

The plumbing you don't have to build

The hard parts of tool calling are the same for every tool. Smarter does them once, for all of them.

The tool definition

Smarter writes the function definition the model reads, from your parameters’ names, types, enums and descriptions. No JSON Schema by hand.

Offered only when it fits

A selector offers a tool always, or only when a prompt mentions its search terms. Fewer tools per prompt means fewer tokens and fewer wrong choices.

Credentials kept out of sight

A manifest names a Connection or a Secret. Passwords and api keys never appear in the YAML, in git, or in what the model reads.

Values, never code

The model only ever chooses parameter values, and Smarter inserts them as data: escaped literals in SQL, query parameters in a request. A clever question can’t rewrite the query.

Cached results

Remote calls are cached, for as long as each plugin says, so the same question doesn’t hit your database or a paid search api twice.

On the record

Each tool call is recorded with the prompt that caused it, inside the assistant’s guardrails and budget, like everything else in Smarter.

See it work

Talk to an assistant whose only tool is a plugin

Stackademy is a fictional code academy with two course advisors. One answers from its database with a SqlPlugin, the other from its web API with an ApiPlugin. The case study lets you ask them both, read their manifests, and change them in the editor.

Every new Smarter platform loads these examples, so they are in your web console to copy from on the first day.

Need a tool no plugin covers?

Write it in Python as an extension, or connect any MCP server. Both are governed the same way as plugins: listed in an assistant, inside its guardrails and budget, and on the record.