Lists redacted providers with their logos and saves records for every type in the provider registry (OpenAI, Anthropic, Google, OpenRouter, DeepSeek, Groq, xAI, Ollama, custom). Two auth modes: an API key, saved through signed SetSecret then SetModelProvider (the dialog refuses to send a key to a remote plain-HTTP server), or Sign in with ChatGPT, offered only when the provider supports it and the server's health reports subscriptionAuth: true, because the flow needs the desktop app to catch the provider's localhost redirect. A provider whose sign-in expired shows a re-sign-in button. ollama and custom get an editable base URL and an optional key. Details in model providers.

Choose server and provider, name the agent, pick a model from the provider's model list, pick a reasoning level where the model supports one, pick a thoroughness preset, and write the system prompt in the Seed block editor. The default prompt embeds the Agent Guide at hm://hyper.media/agent/guide, which the server resolves to its configured knowledge base. Whoever can edit that page shapes every new agent's prompt (prompt injection map). New agents get the default grants: search, web_search, execute, and publish. The dialog first creates a dedicated signing key for the agent, and the server auto-creates an enabled mention trigger following that identity, so mentioning the agent's account starts a session immediately.

Four toggle rows configure the grant set. The verbs and authored tools are not grants:

Toggles autosave. Stored definitions are normalized on read (execute_code becomes execute, the old write-group names become publish, names absorbed into verbs are dropped), so the tab shows what the server acts on. Availability comes from the server's health response: web_search greys out when no search backend is configured, execute when the server reports codeExec: false, with targeted help when the cause is fixable locally. Each row's info button opens the exact model-facing description and schemas from the shared registry. The query and attributes callables have no toggle. See the roadmap.

With Publish enabled the row expands with Author as: the signing identities the agent may sign with, each with its profile icon and name, plus Grant (an existing server identity) and New Account (generates a key, publishes its profile, and creates a home document saying it is an agentic account). Identity management is owner-only, enforced by the server as well as hidden in the UI.

Custom tools continue the list: one row per authored lambda tool document with its runtime badge and summary, an editor for every document field (name, summary, description, runtime, source, input and output schemas), atomic renames, and a destructive delete. MCP servers close the tab: one row per account server with a checkbox that enables it for this agent, a tool-count or Unreachable chip, an Auth chip, and refresh and remove buttons. Clicking a row lists its tools and clicking a tool opens its contract. Add server takes a URL (the name suggests itself from the host), an optional Authorization value stored as an encrypted secret, and transport and header name under Advanced; Connect saves and discovers in one request. See MCP servers.

Durable events from the Log become rows, and finished run cards are interleaved into the same chronological scroll. A #event=<id> hash scrolls to that message. Every row knows its actor, and the actor is checked before the role. User messages are blue bubbles with the sender's live account icon. System messages are quiet grey rows behind a left rule, so continuation prompts and unmet-obligation notices read as runtime notes about the conversation. Agent messages render as assistant parts. A tool row run by a person carries a You chip. Every row can explain itself. A message's info dialog shows the exact markdown the model sees, a share URL, ids, and any hidden window context. A tool row's dialog shows raw input and output. Both end in a Details grid with the sender's account and signer, or the model, provider, duration, and token breakdown.

Tool rows get a purpose-built view per tool: delegate shows the brief and the child's work, execute shows its one-line description, the code, and a live output tail, read and write show the resolved target, a Hypermedia write gets its own phrasing, and a call row borrows the called tool's icon and label. A trigger-created session hides the raw <trigger_context> text and renders a context card instead. Other behaviours: optimistic user messages, concurrent sends while the agent is busy (queued and serialized server-side), live streaming partials, a scroll-to-latest pill, in-app hm:// links, Stop with recovery for sessions stuck in streaming, and Retry only on a trailing error.

The composer is the Seed block editor. Dropped files upload as session-private attachments and are referenced by id. They are never written to memory or IPFS unless the agent does that itself. When a child session is being driven by its parent, the composer is replaced by a line saying so. Beside the send button is the wrench, the person's side of the symmetric log: it lists Read, Write, and every callable the agent is granted and the server can run, with a form generated from each tool's input schema. Running one sends InvokeSessionTool, and the call and result land on the log as user events the agent reads next turn. A contract miss keeps the form open and points at the contract row now in the thread. The wrench is disabled while the agent is busy.

A pinned card sits between the log and the composer while a run is still changing: the run title, a status pill, an elapsed timer, rolled-up token usage, cancel (cascading to the subtree), a context meter for the session, the plan steps with their status icons, children integrated with their steps (one child on a step makes the step that child's row; a batch renders every child as a uniform peer), collapsible tool calls, and two drawers: Code for a script's source and Activity, a terminal-style tail of the whole tree's journal. Once the run reaches a terminal status, a typed result is delivered, or the plan settles, the same card freezes into the transcript and the pinned slot clears. A waiting run never freezes, because the place to answer it is above the composer: children show "Waiting on N sub-sessions", a timer shows a countdown, an event wait renders the run's answerWith signal as an Answer button (with an optional payload editor), and a budget pause renders Resume. Everything on the card reconstructs from ListRuns and the signed runs/<rootRunId> subscription, so a reload loses nothing.

When an agent calls continue_session, the successor opens with a handoff card and a link back to the predecessor. The predecessor shows where the conversation went. The session continuation page has the model.