A real composer
Photos, drawings and dictation; slash commands and @-mentions with live suggestions attached to the field; the model and reasoning pickers the rest of the app uses; interrupt mid-answer.
Cadu
Messenger to your machine.
A native iOS client for Hermes Agent. Every conversation, bot, scheduled task, skill, tool and log line your install runs — carried in your pocket, built foundation-first in Swift 6.
Drag the staff to spin it
The whole install, natively
Not a web view in a shell.
The dashboard, rebuilt for a phone.
Cadu speaks the Hermes gateway directly — REST for the cold facts, a supervised WebSocket for the live ones. Twenty-odd sections of the desktop dashboard were taken apart and put back together in the language a phone actually speaks: lists, sheets, swipe actions, a toolbar menu, Face ID.
Conversations
Every agent's chat,
in one inbox.
The session list fans a listing out across every profile at once and merges the results, so a row carries the agent it belongs to and every write is aimed back at that agent. Pinned conversations lead any order; a job the scheduler started wears a Cron mark.
- Filter and sort that stick. All, active, pinned or scheduled; by project; by any set of profiles — then recent, oldest, title, longest or agent. Which agents you work out of is closer to a setting than a gesture, so it's saved.
- A turn is segments, not one bubble. Commentary emitted alongside a tool call stays where it was made, instead of collapsing into the reply above it.
- Tool calls survive a reopen. Stored history returns tool calls as rows without text; Cadu decodes both shapes and puts the sequence back together, arguments and all.
- Approvals and clarifications inline. When the agent asks permission or asks a question, the answer is a control in the transcript — not a paragraph you have to type.
Everything the reply carries
Generated images resolved from MEDIA: references, markdown or bare artifact paths, then fetched and swapped in as real attachments. Source files open in a syntax-highlighted viewer; anything else in QuickLook.
Read it back later
Reactions, threads, a session inspector, presentation-only display controls, per-conversation artifacts — and a SwiftData read-through cache so a thread opens before the socket does.
Bot Mode
A team roster,
not a config file.
Hermes Agent profiles rendered the way a messaging app renders people. Presence, recency,
custom names, geometric faces or uploaded and generated portraits, groups, and a preview of
the canonical chat — all read from profiles.list and profile assets, so every
client agrees on who's who.
One forever-chat each
Every bot opens its own server-pinned, hidden Bot Chat. A new bot introduces itself automatically, and /new compacts that chat instead of forking a fresh one.
Group rooms
Bounded two-to-six member rooms assembled by the app out of ordinary profile-scoped sessions, with their own lavender identity so a group never reads as a one-to-one.
Validated handoffs
Mention another bot and the turn is handed over — checked against the roster first, so a typo is a correction rather than a silently dropped message.
Full profile editing
Create, duplicate, delete; edit capabilities, skills and portraits; browse a bot's own sessions and its bot-scoped routines. Every write stays profile-scoped through the gateway.
Automation
What the agent does
while you're not looking.
- Tasks. Every scheduled job across every profile, because a schedule is an install-wide thing to watch. Pause, resume, run now, edit and delete are swipe actions; a row opens the configuration and the runs it has already produced — the question worth asking about a cron job is whether it worked.
- Four schedule grammars, one picker. The editor reads and writes what the server accepts, and an expression it can't represent survives untouched rather than being approximated into a different schedule.
- Blueprints. The server's catalog of ready-made automations, each form built from the field schema the blueprint ships — so one added server-side appears without an app release.
Skills
A switch per skill, filtered by state and category. Toggling is optimistic and puts itself back if the server refuses. A row opens its SKILL.md in an editor; the server validates the frontmatter and its complaint is what you read.
Toolsets
The tool groups Hermes exposes to one profile, switched natively through the shared management API.
Plugins
Install by the same repository identifiers and Git URLs Hermes takes. Rows switch optimistically, expose updates only for Git-backed plugins, and permit removal only when the hub marks it safe.
MCP
Servers and the catalog they come from, each row showing in full what it actually connects to or runs. Enable, test and remove by swipe; a refusal is a result, not an error. OAuth is completed by the server — nothing to register on device.
Insight
What it cost,
and what it said.
Token, cost, model, skill and tool accounting straight from the profile database — with the auxiliary work most dashboards hide, like compression and vision passes, named and priced alongside the models you chose.
- Logs that read like records. The API returns physical lines and only the first carries a level, so a traceback would arrive as four unlabelled lines under the error that raised it. Cadu folds them back into one record and colours the block.
- Filters behind one button. A phone has room for the log or for four segmented controls — so minimum level, component and free text live behind a toolbar button, with what they're set to spelled out above the lines.
- System. Gateway state, session and agent counts, restart and update behind confirmations — and nothing invented that
/api/statusdidn't confirm.
Everything else
The rest of the app,
at a glance.
Sessions
Stored conversations across every agent, merged, filtered, sorted and pinned.
Bots
The roster, their forever-chats, groups and handoffs.
Live voice
A hands-free conversation transcribed and spoken on device: silence sends, and speaking over the answer interrupts it.
Projects
Start a conversation rooted at a path on the connected host.
Artifacts
The durable outputs a conversation produced — images, files and links — mined from assistant and tool messages.
Files
Hermes' managed files: listing, drilling in, sorting, preview, delete and new folder.
Models
What new sessions will run and what recent ones did, through the same picker the composer uses.
Usage
Tokens, cost, sessions and API calls by model, skill, tool and day.
Logs
A tail of any of the six log files, by level, component and text.
Tasks
Cron jobs and blueprints across every profile, with their run history.
Skills
Per-skill switches and a validating SKILL.md editor.
Toolsets
Profile-scoped tool groups.
Plugins
Installed plugins, their state, source, version and updates.
MCP
Servers, catalog, probes, installs and the OAuth handshake.
Messaging
Pending users with platform, identity and age; approve or revoke the whitelist.
Profiles
Create, edit and switch agents, with avatars you can upload, generate or leave geometric.
System
Health, activity and maintenance for the connected install.
Kanban
The Kanban plugin's board, each lane a horizontally paged card instead of seven squeezed columns.
Achievements
The badges plugin as a collection hero, compact stat grid, tier legend and one readable card per badge.
Push
One button installs the relay plugin and hook into every profile home, pairs the device and bounces the gateway.
Instances
Any number of installs, address, account and password together in the Keychain.
App lock
Face ID in front of everything, with a grace period and a passcode fallback.
Onboarding
What Hermes is, how it should look, and the commands that prove the install is reachable.
Questions
Before you install it.
Do I need my own Hermes Agent install?
Yes. Cadu is a client, not a service — it talks to a Hermes Agent installation you run. Point it at the address, and it probes /api/status before asking for a password, so an unreachable host reports as unreachable rather than as a failed login.
Can it talk to more than one install?
As many as you like. Each saved instance carries its address, its account and — if you asked it to — its password, together as one Keychain item. Tapping a saved instance fills the form in and connects.
Does anything go through a server of yours?
Only push notifications, and only if you switch them on: the relay knows which devices an install key fans out to, and the notification itself is posted by your own install's adapter. Everything else is a direct connection between the app and your host.
What happens when the network drops?
Retryable failures reconnect with exponential backoff and full jitter, and the stored session resumes — a train tunnel costs you a pause, not a conversation. Transcripts are cached on device, so a thread opens before the socket does.
Is it usable on an iPad?
Yes. The same sections run behind a sidebar on a regular width and behind a three-slot tab bar on a compact one, and which three sections those are is yours to choose.
Messenger to your machine.
Cadu is in open beta on TestFlight. Point it at a Hermes Agent install and it starts talking — the address is probed before a password is ever asked for, so an unreachable host reports as itself.