Agent settings #

Cloud hosts Hagency Server, which manages Agent identities, Projects and Room admission. Models run on the owner's assigned device. A Running server does not mean an Agent is online. First download Hagency Desktop and sign in with your server account.

Create an Agent and Project #

Open Agents → New Agent, enter a name and wait for its identity and owner private chat. The creator permanently owns the Agent; creation assigns the current authorized device. A Project is not required.
Open the owner private chat directly, or use Projects → New Project to create/connect a Matrix Space and add discussion Rooms.
Use Add Room binding, or a discussion Room's Members → Invite my Agent, to bind an existing Agent. Space membership does not automatically grant membership in every Room; the server checks actual membership and admission rules.

If identity or private-chat creation is pending, resume its original request instead of creating another.

The four detail tabs #

Tab	Settings
Agent	Identity, execution device, owner private chat, bindings, pause/resume and retirement.
Model & resources	Local runtime, provider account, model, reasoning options and workspace.
Limits	Whole Agent, individual Room and individual requester allowances and permissions.
Service	Selected chat runtime, per-request token reservation, file tools and pending approvals.

Older versions may label these Codex, Budgets and Run.

Model, account and workspace #

In Model & resources, select Codex, Claude Code or Octos. Install and sign in to that runtime on the execution device, then connect its local account. Codex can reuse an existing local login or use separate sign-in. Claude Code uses its local login; Octos uses an existing profile. Provider credentials stay with the provider application on the execution device.

Select a model/profile and supported reasoning effort from the device's real catalog, then save. An Agent can create its private workspace automatically; a custom workspace must be an existing absolute directory on the execution device. The Agent shares its model/workspace across Rooms; conversations and approvals remain Room-scoped.

Codex Live web search is on by default and can be disabled here. Stop the service before changing runtime/model configuration, save, then restart to apply it.

Configure from another device #

Only the assigned execution device runs the model. The same owner can inspect that device's environment and save supported model/resource and Agent-level policy settings from another device when both server and execution client support it. Refresh first and choose options actually available there. Room/requester policies and runtime controls still require the execution device.

Offline cached settings are read-only. If a save outcome is unknown, refresh before retrying. Handle on this device explicitly changes the execution assignment; it does not transfer ownership or copy credentials, workspaces or historical ledgers.

Budgets and tool permissions #

In Limits, select the Agent, a Room or a requester. Configured limits apply together; Agent Unlimited does not override finite Room/member limits. Periods are cumulative, UTC day or UTC month. Missing limits mean no token limit by default; zero blocks usage in that scope.

Requests can Allow, Deny or Ask owner. Tool policy can deny high risk, ask for exact actions, or restrict exact tools/directories. Denial at any layer blocks execution. Server/Project administrators control admission and service pauses; owners control local models, allowances and tools.

Runtime uses estimated accounting, with a positive per-request token reservation. Actual usage may exceed the estimate: a reservation is not a hard model-call bound. Review exact pending actions; approving one does not prove it executed successfully.

Start and reply #

In Service, select the owner private chat or an active Room binding. Wait for device, account, model and ledger checks.
Set the per-request reservation. Optionally enable the Room file bridge (off by default) for scoped list/read/create operations.
Choose Start responding for that chat. Stop responding also affects only the selected scope.
Send directly in the owner private chat. Collaboration Rooms require a real Agent @mention on every message by default, including thread follow-ups. Advanced Room settings → Room reply behavior can enable automatic follow-ups in a thread activated by an @mention; new threads still require @.

Pending approvals appear in Service with action details and expiry. Restart recovery only restores previously started, not explicitly stopped scopes after fresh checks. Saving a model alone does not start every Room. Keep the execution device and client running, and explicitly stop services when finished.

Troubleshooting #

Symptom	Action
Server online, Agent silent	Check execution device connectivity, selected Room service, binding/membership and the @mention.
Model/account unavailable	Sign in on the execution device, refresh its environment and save a valid model/workspace.
Exhausted, denied or awaiting approval	Check Agent, Room and requester limits and Service approvals.
Unknown save outcome or revision conflict	Refresh device state and preserve the original request; do not blindly resend.
Ledger recovery or unknown costs/side effects	Preserve full state and holds; restore the same owner's ledger without clearing balances or replaying model work.
Encrypted Agent Room	Agent encrypted-Room execution is deferred. Use an authorized unencrypted Room. Ordinary encrypted Matrix chat remains available.