The kapa R package enables creating ellmer tools from kapa.ai projects. kapa.ai projects compile context from various resources and surface that information through search.
You can install the development version of kapa like so:
pak::pak("simonpcouch/kapa")Once the package is installed, you will need a kapa.ai API
key. Set this to
the KAPA_API_KEY environment variable.
kapa_search_tool() returns an ellmer tool for the provided kapa.ai
project. kapa_search_tool(project) defaults to
Sys.getenv("KAPA_PROJECT_ID")—I’ve set mine to a
project that compiles documentation
and GitHub history on a few tidyverse packages:
library(kapa)
tidy_search <- kapa_search_tool()
tidy_search
#> # <ellmer::ToolDef> Posit_kapa_search(query, num_results, include_source_names)
#> # @name: Posit_kapa_search
#> # @description: Search the Posit knowledge base with sources Corrected
#> Responses, dplyr, ggplot2, and tidyr.
#> # @convert: TRUE
#> #
#> function (query, num_results = 10L, include_source_names = NULL)
#> {
#> kapa_search(query = query, project_id = project_id, num_results = num_results,
#> include_source_names = include_source_names)
#> }
#> <bytecode: 0x1158eb970>
#> <environment: 0x1158ea668>I can add this tool to any ellmer chat object to allow the chat to
search this kapa.ai project before responding to questions. For example,
coord_flip() was deprecated after Claude Sonnet 4 was trained, but
using the kapa.ai search allows the model to discover that the function
was deprecated before responding to the user:
library(ellmer)
client <- chat_anthropic(
"Use the provided tool to research your solution before responding tersely.",
model = "claude-sonnet-4-20250514"
)
client$register_tool(tidy_search)
client$chat(
"
Could you flip the axes in this plotting code?
ggplot(diamonds, aes(x = carat)) +
geom_histogram() +
scale_x_reverse()
"
)
#> ◯ [tool call] Posit_kapa_search(query = "ggplot flip axes coord_flip
#> histogram x y axis")
#> ● #> {"search_results":[{"content":"# ggplot2 > Reference\n## Cartesian …
#> To flip the axes in your plotting code, you have two main options:
#>
#> **Option 1: Using `coord_flip()` (keep your current code structure)**
#> ```r
#> ggplot(diamonds, aes(x = carat)) +
#> geom_histogram() +
#> coord_flip() +
#> scale_x_reverse()
#> ```
#>
#> **Option 2: Swap the aesthetics (preferred modern approach)**
#> ```r
#> ggplot(diamonds, aes(y = carat)) +
#> geom_histogram() +
#> scale_y_reverse()
#> ```
#>
#> The second option is preferred because `coord_flip()` is now superseded. When
#> you map the variable to the `y` aesthetic instead of `x`, the histogram will
#> automatically be oriented horizontally, and you use `scale_y_reverse()`
#> instead of `scale_x_reverse()` to maintain the reverse scaling.To start an MCP server with this tool, use
mcptools::mcp_server() with
tools = list(<your_tool>).