Tools

Every chat session starts with the built-in toolset, plus whatever MCP servers and local plugin tools (~/.config/kaja/tools/*.ts) you’ve added.

Tool	Purpose
`read_file` / `list_files`	read local files, list a directory
`fetch_url`	fetch a URL
`view_image`	look at an image file
`summarize`	summarize long text
`rerank`	rerank passages against a query
`current_time`	current date/time
`ask_user`	ask a clarifying question mid-task
`run_command`	run a shell command
`switch_persona`	change persona mid-conversation
`remember_note` / `recall_memory` / `forget_note` / `list_notes`	long-term memory notes
`dataset_info`	collect structured answers for a persona’s dataset

Two more appear when configured: web_search (needs [webSearch] in secrets.toml) and generate_image (needs a models.image-generation entry).

Shell commands

run_command always asks before running. Known-risky patterns (rm -rf, sudo, git push --force, DROP TABLE, fork bombs, …) get a louder warning — an advisory cue, not a sandbox. It runs with your own shell permissions.

MCP servers

Add a server to ~/.config/kaja/mcp.toml and its tools are folded in automatically — local (stdio, needs command) or remote (HTTP, needs url):

[[servers]]
id = "playwright"
command = "bunx"
args = ["@playwright/mcp@latest", "--isolated", "--headless"]

[[servers]]
id = "geo-service"
url = "https://your-geo-service-host/mcp"

Needs a header or env var with a secret in it? Add it to secrets.toml under [mcp.<id>], keyed by that server’s id:

[mcp.geo-service]
Authorization = "Bearer your-secret-api-key"

Playwright ships enabled by default, giving the agent a headless browser.

Your own tools

Drop a .ts file exporting a tool under ~/.config/kaja/tools/ — picked up automatically, no rebuild needed.