AI agent
Agent tool calls
The agent picks tools, passes args, reads results, decides next steps — visible in the chat stream.
How tool calls actually work
When you send a prompt, the agent ships every available tool schema to the model. The model can answer directly or "call a tool first." After the result comes back, the loop continues — sometimes for many rounds.
Kition renders each round as a collapsible card: which tool, what arguments, what it returned, how many tokens. This is the difference between an agent and a black box.
Built-in tools
Read/Write/Edit— file ops, scoped by path globBash— shell execution, defaults to vault rootweb_search/web_fetch— web search and fetchbrowser_open— real Chromium browser automationtable_query/table_upsert/table_delete— table opsgrep/glob— cross-file search- MCP — third-party extensions (GitHub, Jira, custom tools)
What a real call looks like
You say: "Find docs with TODOs from the last 7 days and list them in a table." The agent typically greps, then reads, then table_upserts, then sends a confirmation. Every step is expandable in the panel.
> tool: grep
pattern: "TODO|FIXME"
path: "**/*.md"
result: 14 matches across 6 files
> tool: Read
path: "projects/onboarding.md"
result: 2.3 KB
> tool: table_upsert
table: "todos"
rows: [
{ file: "projects/onboarding.md", text: "wire up SSO", owner: "?", due: "?" },
...
]
result: 14 rows insertedSee every step, edit any step
- Click retry — rerun with the same args
- Click edit args — adjust and run again
- Click skip — fake success and let the agent continue
- Click stop — halt the whole loop
Parallel tool calls
When tool calls have no dependencies, the agent dispatches them in parallel — e.g. reading five files at once. This is the main speed gap between Kition and "one-step-at-a-time" agents.
Tune the fan-out in Settings → Agent → Concurrency: lower it to be polite to providers, raise it to pin your local CPU.