Kagi Assistant

Kagi Assistant

Use Assistant for conversation mechanics and continuity. Use kagi-ai to choose when to spend AI allowance, and kagi-usage for direct search or page work.

Assistant commands require KAGI_SESSION_TOKEN.

Start a conversation

kagi auth status
kagi assistant "draft a migration checklist" --format markdown
kagi assistant "compare these options" --format toon

State the deliverable, constraints, and audience in the prompt. Never rebuild the output structure in follow-up shell processing when the Assistant can produce it directly.

Continue a thread

kagi assistant --thread-id THREAD_ID "add rollback steps"
kagi assistant thread list
kagi assistant thread get THREAD_ID
kagi assistant thread export THREAD_ID --format markdown

Reuse a thread only when its prior context still belongs to the task. Start a new thread when stale context could bias the answer.

Delete a thread only when the user asks for it:

kagi assistant thread delete THREAD_ID

Attach local context

kagi assistant \
  --attach ./notes.md \
  "turn these notes into a decision record" \
  --format markdown

Attach only files the task needs. Name the file's role in the prompt so the Assistant knows whether it is evidence, a template, or background.

Stream responses

Use human streaming for interactive terminal work. Use structured streaming when another program consumes events.

kagi assistant --stream "explain the tradeoffs"
kagi assistant \
  --stream \
  --stream-output json \
  "produce a release checklist"

Never parse human streaming output as a machine contract.

Custom assistants

kagi assistant custom list
kagi assistant custom get "Researcher"
kagi assistant custom create \
  "CLI Researcher" \
  --web-access \
  --model gpt-5-mini

Inspect an existing custom assistant before changing the workflow around it. Account configuration belongs in kagi-account-config.

Completion criteria

Assistant work is complete when:

• you used the correct thread or a clean new one;
• you supplied the required files and constraints;
• the output format matches its consumer;
• the final response answers the current prompt, not stale thread context; and
• you deleted a thread only when the user asked for it.