Skip to main content

Feature Overview

This page maps AgentHub's main capabilities to the problems they solve.

Single-Agent Control Surface​

From the Agents page you can:

  • create, start, stop, reconnect, and delete agents
  • choose use_existing or create_worktree
  • point execution at the main node or a remote node
  • keep one long-lived session instead of restarting from scratch after every browser refresh

Use this path when one person is directly driving one coding or automation session.

Structured ACP Output​

AgentHub renders ACP sessions as structured timelines instead of plain terminal logs. That gives you:

  • conversation history
  • tool-call and command output context
  • plan and reasoning segments
  • debug and raw-event inspection when something looks wrong

Use this when task review, auditability, or deep debugging matters.

ACP debug surfaces can also expose runtime-controlled session actions, such as:

  • provider-driven mode / model selectors
  • generic session config updates
  • force-new-session recovery when a thread gets stuck

The exact controls depend on the active ACP provider and runtime.

Team Workbench​

/teams adds a multi-agent workspace surface with clear role boundaries.

Key concepts:

  • Channels lead the Team shell, with # all as the default shared lane
  • the shared channel is stable Team state, not inferred from the visible Kanban slice
  • Kanban is the canonical task lane
  • Execution Runs and debug surfaces are execution artifacts, not the primary planning surface
  • Agent ACP lets you inspect a specific member deeply when needed

Use this when planning, implementation, and review should be coordinated rather than collapsed into one session.

Recent Product Updates​

Recent merges worth knowing before rollout or operator training:

  • Workspace shell convergence: Team pages now use a more compact workspace shell with clearer Channels-first navigation and less duplicated chrome.
  • Execution wording cleanup: Team run history is presented as Execution Runs to distinguish planning state from raw execution artifacts.
  • Token-based Agent Node onboarding: remote node join now uses explicit bootstrap tokens from the Agents page instead of QR-style node onboarding.
  • Rust / Codex baseline refresh: local builds and bundled ACP integration now track Rust 1.95.0 and the official Codex 0.121.x line.

Installable Web App And Push​

AgentHub can also behave like an installable web app:

  • supported browsers can install it as a standalone app window
  • the frontend registers a service worker automatically
  • browser push notifications can be attached to the signed-in browser session

AgentHub still prefers fresh server-delivered UI resources on reload; install support does not turn it into an offline-first shell.

Agent Nodes And Actor P2P​

AgentHub can register remote execution nodes and route agent execution to them over internal gRPC, while keeping actor control and mailbox delivery on the same p2p-aware control path.

Use this when:

  • repositories or compute live on other machines
  • you want one main control plane but many execution targets
  • node-specific default worktree roots are required
  • actor-to-actor coordination should still work even when execution leaves the main machine

Automation And Integration​

AgentHub exposes authenticated OpenAPI discovery endpoints for:

  • internal tooling
  • typed client generation
  • contract checks
  • operational automation around agents and Teams

Use this when AgentHub needs to plug into a larger engineering system.

Operations And Safety​

AgentHub also includes operational guardrails:

  • safe-path restrictions for workdirs
  • persistent session and history storage
  • in-app notifications and browser push delivery
  • troubleshooting and recovery guidance

Next Reading​