ARIA & the DAEMON Console

The chat-first operator in the right rail that drives the whole IDE from natural language.

Overview#

The DAEMON Console lives in the right rail (toggle it with Ctrl+B, or move it to the bottom panel beside the terminal). Inside it runs ARIA, an operator agent that can run DAEMON itself: read your wallet, preflight and launch tokens, commit changes, scaffold files, change settings, and fan out swarms of parallel agents — all from chat. Type / for commands like /wallet balance, /solana cluster, and /launch tokens, which answer inline in the thread.

Per-Project Sessions#

Every project gets its own chat sessions. Create, switch, rename, archive, or delete them — titles are generated automatically from your first message, and conversations rehydrate from the local database when you restart the app.

Compounding Memory#

ARIA records durable project facts as you work — decisions, build and test commands, constraints, prior failures and fixes, deployment targets — and recalls them in later sessions. Memory is stored locally, can be listed, corrected, or archived, and makes the operator more useful the longer you work in a project.

The Tool Catalog#

ARIA works through a registry of 40 typed tools spanning the whole IDE:

DomainWhat ARIA can do
NavigationOpen panels and files, run command-palette commands, read project status
WalletRead balances, generate and assign wallets, store RPC keys
Token launchPreflight cost estimates, list launchpads, deploy tokens
GitStage and commit changes (it never pushes)
MemoryRemember, recall, update, and archive project facts
AgentsManage Clawpump and AgentStation agents, scaffold agent projects
SwarmsLaunch, monitor, and collect parallel agent lanes
WorkspaceScaffold files and directories, run engine health checks
SettingsChange app settings, enable or disable integrations

Plan Mode vs Build Mode#

  • Build mode (default) — ARIA acts on your request immediately
  • Plan mode — ARIA presents an ordered plan first and waits for one approval before writing anything

Safety Model#

  • Read tools auto-run — navigation, wallet reads, and memory recall execute without prompts
  • Write tools show approval cards — commits, file scaffolds, and memory writes wait for an inline approval
  • Sensitive actions need typed confirmation — wallet generation, token launches, and swarm launches require you to type the confirmation phrase
  • [MAINNET] guard — when the cluster is mainnet-beta, approval cards and tool results carry an explicit marker so real-money actions are unmistakable
  • No autonomous pushes — ARIA can commit, but pushing to a remote is always yours

Agent Swarms#

Ask ARIA to fan a job out and it launches 2–12 parallel tasks, each in its own isolated git worktree and branch, each driven by an independent Claude lane (up to 4 run concurrently). Every lane writes a RESULTS.md summary at its worktree root; worktrees are cleaned up automatically, and merging lane branches is a deliberate manual step in the Git panel.

Swarms differ from Grind Mode: grind panels share your working tree for interactive parallel work, while swarm lanes are fully isolated worktrees you review and merge afterward.

Model Lanes#

On the free Light tier, ARIA runs through your own Claude setup (BYOK) with three lanes: fast (Haiku), standard (Sonnet, default), and reasoning (Opus). Pro and above route through DAEMON AI Cloud with five hosted lanes — auto, fast, standard, reasoning (Operator+), and premium (Ultra). Pick the lane from the model dropdown in the Console header.