Your agent
Your agent is the one you talk to. It knows you, keeps your memory and your tasks, and hands code work to squads. It runs on one of your hosts, so it does not depend on your computer being on.
What it is made of
Section titled “What it is made of”| Part | What it is |
|---|---|
| Name | Lower case letters, digits, ., _, -. The first one is called kedalio. |
| Engine and account | Claude Code, Codex, OpenCode or Pi, and one of your engine logins (Engines and costs). |
| Model | Optional: the engine’s default otherwise. |
| Rules | Its prompt: how you work, what it may do alone. Up to 20,000 characters. |
| Host | Optional: where it runs. Without one, any host online with room. |
| Delegation | What it may do without asking you. By default: Kedalio’s own tools only. |
Your agent is made the first time it is needed, with the only engine login you have if you have exactly one. Every change is kept as a revision and can be undone. In the app it is System → My agent: the rules and their revisions, the engine, account, model and host, how long it stays on, its tools, and the changes squads propose.
Every run of your agent also gets your active skills and Kedalio’s own tools: memory and tasks, squads (it proposes briefs), skills. It does not add projects or assignments: you add those in the app (Projects) or from Claude or ChatGPT with the connector (kedalio.project_set, tasks.assignment_add).
What else it may use is yours to switch on, in My agent → Tools: web reading and web search, and the MCP servers you added (Tools and integrations). Its tools ask you first unless you allow them.
In your own Claude Code sessions with the Kedalio plugin, the integrations’ tools are native tools (mcp__kedalio__<name>), see the connector. Your agent’s runs do not have them yet: they reach those tools through the tools door.
Talking to it
Section titled “Talking to it”In the app, open Chat: your conversations are on the left, and you can rename or delete each. While the agent works you see the tools it uses, and when it needs a permission the request is in the conversation, with Allow and Deny. When the work calls for a squad, it proposes one there (below).
A chat is an engine session. The server keeps it, sealed with the vault’s key, and a chat that went quiet picks up where it stopped when you write again: the agent remembers what it read and did, not only what was said. This holds for Claude Code; an agent on another engine starts again from the last ten turns. How long it stays on (15 minutes, 1, 4 or 24 hours, or always; 60 minutes by default) only trades a faster answer against a host slot kept busy.
From Claude or ChatGPT with the connector, the client calls work.chat, which sends your words and waits for the answer (up to 50 seconds). When an answer comes and nobody is following the chat, you get a notification.
To change the agent from a client, ask it to use work.myagent_set (name, engine, account, model, host, delegation) or work.myagent_rules (its rules); work.myagent_get shows the definition.
Getting to know you
Section titled “Getting to know you”The first conversation is an onboarding: the agent asks, in order, who you are, how you work, your projects and your accounts. It writes what it learns into your memory and into its own rules, can read a ChatGPT or Claude memory export you paste, and suggests a short list of starter skills. Onboarding is a step before the app, with a «Skip» button always in view; it is skipped on its own when your memory already has pages about you or any task.
The agent keeps improving its rules as you correct it, with work.myagent_rules.
From a request to a brief
Section titled “From a request to a brief”Your agent does not edit repositories itself: its run has no repository. When you ask for code work, it acts as a prompting master and writes a brief for a squad, with five parts: Goal, Context (from your memory), Done (what «done» means), Verify (how to check it), Budget.
The brief is proposed to you in the chat and with a notification («a brief to approve»), and it also waits in Today → To decide. Its page is where you approve it, edit it first, or drop it; from Claude or ChatGPT the same is work.briefs and work.decide. Approving starts the squad. Your last ten edits are shown to the agent with every brief it writes, so it learns how you want them.
Your agent never answers its squads’ requests on your behalf: those reach you (Squads and assignments).
Skills
Section titled “Skills”Skills are folders with a SKILL.md, the open format Claude Code, Codex and OpenCode read. Your library is on the server; a run gets up to 40 active skills. A squad may write a trial skill for its own project; you promote it to the project or to your whole account. After a squad that went badly, a retrospective proposes changes to skills, which you approve, edit or drop. All of it is in Skills.
Kedalio ships its own skills for routing work across models (Engines and costs). A catalogue of outside skills is not built yet.
Not built yet: suggestions you install with one tap, and voice.