What your agent can do with kcap

Setup registered a set of kcap tools in each agent you chose, plus a few skills: instruction documents the agent loads when a request matches. You never call these by name. You ask in plain language and the agent picks the tool

Each section below is one kind of thing you can ask, with a few example prompts. The examples are not commands to copy; they show the shape of a request, and anything you would say in the same spirit, in your own words, works the same way

Remember what the team did

What have we done on rate limiting in this repo? Find the sessions and summarise them

Recap session acme-billing-rate-limit and tell me what is left

Who else is working in this repository right now?

The agent searches sessions you are allowed to see, reads their summaries and transcripts, and cites them. Ranking favours exact words, so an issue key, an error message or a function name finds the right session faster than a description. Each result says whether it matched literally or only by meaning

Continue work

Is there an unfinished plan here? Continue from its next task

Continue the work from session 3f2a…; it was interrupted

The agent finds open plans for the repository, attaches this session to one, and carries the task list on. Continuing an interrupted session also attaches its work items. If kcap thinks the other session may still be running, the agent is told and asks you before forcing it

Keep a plan on the record

Declare the plan for this task and keep the tasks updated as you work

The agent records its plan document and task list, and updates each task as it starts, finishes or is skipped. The session’s Overview shows the list live; a later session can resume it

File the session on a ticket

This session is for AI-1234. Attach it

Break AI-1234 into these two parts and record that the second is blocked by the first

An explicit declaration beats automatic correlation and sets the work item’s primary issue. Parts and dependencies are recorded on the work item, where they show as cards

Decide what is next, and leave loose ends

What should I work on next?

I’m stopping here; leave a loose end for the migration script

Claim that loose end and close it when done

The next-work feed is ranked and comes with reasons. A loose end is a concrete, claimable piece of unfinished work; a claim stops two people doing it

Save and use team memory

Remember for the team that this repo uses pnpm. Make it repository-wide and visible to everyone

What do we already know about deploying this service?

Archive the old staging memory; it no longer applies

Say where a memory applies (repository, project, organisation) and who may see it (you, a team, everyone). At session start the agent gets an index of relevant memories, not their bodies, and reads one when it needs it

Curate knowledge

Show me the facts behind the “always run integration tests” skill and correct the one about Postgres

Move that skill out of review and into serving

Agents can read facts and skills, fix a fact’s wording, edit a skill body and move a skill along its lifecycle. The web app’s Knowledge pages do the same with more context

Review before merging

Start a code-review flow on this branch with Codex as the reviewer, and iterate until it is clean

Run the spec-review flow on this design document

A flow launches a separate reviewer agent on a daemon and runs rounds until sign-off. The agent lists the flows your workspace has when you ask

Publish a Page and wait for answers

Write up the three migration options as a Page with a comparison table, share it with the platform team, and wait until two of them have answered before continuing

A Page is a self-contained HTML document at a stable URL with an explicit audience, and can carry a form the agent blocks on

Hide this session, or stop storing it

Hide this session

Stop storing this session and delete it

Hide makes the session owner-only and keeps storing it. Stop ends storage for this session and deletes it from the server

Start another agent

Hand the test-writing to a separate hosted agent on my daemon and carry on with the refactor

The started agent is an ordinary hosted agent, visible under Agents, and this one does not wait for it

Ask about usage and cost

Which repositories used the most tokens last month, and which models?

On Team and Enterprise plans the agent can query the same governed views Insights uses: counts, tokens, cost, tool use and evaluation scores, never transcript text

What the agent cannot do through kcap

It cannot read a session you cannot see, share something wider than you could, or launch a reviewer on a daemon whose owner has not allowed you. Tools that widen an audience or launch paid work (flows, memory, pages) are never pre-approved in agents that have a per-tool trust setting; the agent asks first

The full tool inventory is in MCP servers, tools and skills