AI代理仪表盘

Prerequisites

Key Dependencies

Install

Step 1: Install the binary

Download the pre-built binary from the latest GitHub Release:

curl -fsSL https://raw.githubusercontent.com/bjornjee/agent-dashboard/main/install.sh | sh

The installer downloads the binary for your platform, verifies its SHA256 checksum, and installs it to ~/.local/bin/agent-dashboard. Hooks and skills are delivered through each host's plugin marketplace (see Step 2 and the Codex section below); the installer does not write into ~/.codex. No Go toolchain required.

Or build from source (requires Go 1.26+):

git clone https://github.com/bjornjee/agent-dashboard
cd agent-dashboard
./install.sh --build

Step 2: Register the marketplace and install the plugin

In any Claude Code session, run:

/marketplace add bjornjee/agent-dashboard
/plugin install agent-dashboard@agent-dashboard
/plugin enable agent-dashboard@agent-dashboard

Then restart Claude Code sessions for hooks and skills to take effect.

Uninstall

Usage

Run the dashboard directly:

agent-dashboard

Or if you set up the tmux keybinding, press prefix + D to switch to a dedicated dashboard session.

Optional: tmux keybinding

The included agent-dashboard.tmux script binds prefix + D to switch to a dedicated dashboard session:

```bash

Reload tmux config

tmux source-file ~/.tmux.conf ```

User Settings

The dashboard supports a TOML configuration file at ~/.agent-dashboard/settings.toml (or $AGENT_DASHBOARD_DIR/settings.toml if overridden). From a repo checkout, the installer copies this from settings.example.toml when the destination file does not already exist. Any missing keys fall back to sensible defaults — you only need to include the settings you want to change.

Example settings.toml:

[banner]
show_mascot = true   # show the axolotl pixel art (default: true)
show_quote  = true   # show the daily quote (default: true)

[notifications]
enabled       = false  # enable desktop notifications from adapter hooks (default: false)
sound         = false  # play alert sound on attention events (default: false)
silent_events = false  # show notification for non-alerting stops (default: false)

[debug]
key_log = false       # write key/mouse/focus events to debug-keys.log (default: false)

[experimental]
ascii_pet = false     # show animated ASCII pet in the left panel (default: false)
dino_game = false     # show Chrome-style dino runner game in the left panel (default: false)

[usage]
rate_limit_poll_seconds = 60  # how often to fetch rate limits from Anthropic API (default: 60, 0 = disable)

[effort]
plan    = "high"  # thinking-effort level pinned while permission_mode='plan' (default: high)
default = "high"  # thinking-effort level pinned at spawn and restored on plan exit (default: high)

The [effort] levels feed the /effort slash command Claude Code accepts (low | medium | high | xhigh | max). The agent-state-fast hook swaps in plan when the agent enters plan mode (EnterPlanMode) and restores default on exit. The feature, fix, and refactor skills additionally declare effort: max in their frontmatter, which Claude Code pins for the skill's lifetime when the skill is invoked as a slash command inside an existing session.

Environment Variables

Codex CLI support

Codex support is packaged as a Codex plugin adapter in adapters/codex/. Register the marketplace entry with Codex:

codex plugin marketplace add bjornjee/agent-dashboard

Then enable the plugin by appending the following to ~/.codex/config.toml:

[plugins."agent-dashboard@agent-dashboard"]
enabled = true

Restart Codex sessions and approve the agent-dashboard hooks prompt. Once approved, the dashboard sees Codex sessions just like Claude sessions — same state file, same conversation panel, same cost dashboard. Run codex --model gpt-5.5 in a tmux pane and the agent appears in the dashboard's agent list.

From a repo checkout you can also run make install-codex-adapter, which performs the marketplace registration and prints the config snippet.

Caveats specific to codex:

• Some skills remain Claude-only. /implement and /rca still rely on Claude-only orchestration primitives. The dashboard returns a 400 for unsupported codex skill combinations.
• Plan mode is signaled, not gated. Codex's /plan slash command flips the hook payload's permission_mode to "plan". The dashboard captures this as a field but doesn't flip state to plan — codex has no ExitPlanMode equivalent, so there's no discrete "plan ready" review moment.
• No subagent tree for codex. Codex doesn't have a Task/Agent tool, so the subagent tree panel stays empty for codex sessions.

Recommended companion plugins

agent-dashboard works best when paired with bjornjee/skills — a plugin of workflow skills (TDD guide, language-specific strict reviewers, refactor cleaner, codex delegation, terminal ops) that the dashboard's session-creation flow expects:

/marketplace add bjornjee/skills
/plugin install skills@bjornjee-skills

Without it, skill-gated session types (feature, fix, refactor, pr, rca) will not function as intended.

Step 1: Remove the plugin

In any Claude Code session, run:

/plugin uninstall agent-dashboard@agent-dashboard
/marketplace remove agent-dashboard

FAQ

Do I need tmux? Yes. agent-dashboard reads live pane content via tmux capture-pane and spawns agent sessions in tmux panes. Without tmux there are no panes to monitor.

Which agents are supported? Claude Code is first-class via the adapter in adapters/claude-code/. Codex CLI is supported directly via the adapter in adapters/codex/ (see "How do I see codex agents in the dashboard?" below). Codex is also reachable via skill delegation (/codex-delegate). The architecture supports additional backends via the domain.Harness interface.

How do I use codex / gpt-5.x models? Pick codex in the New Agent harness step (TUI wizard or web form) to spawn the Codex CLI directly, with per-spawn flags from [harness.codex] in ~/.agent-dashboard/settings.toml.

How do I see codex agents in the dashboard? Codex sessions appear once the Codex plugin is installed, enabled, and its hooks approved. Run codex plugin marketplace add bjornjee/agent-dashboard, enable the plugin in ~/.codex/config.toml, restart Codex, and approve the agent-dashboard hooks prompt. See the Codex CLI support section below for details.

Does this require a paid Claude account? No — it uses whatever Claude Code itself requires (Pro, Max, or API). agent-dashboard does not call the Anthropic API directly; it reads the JSONL transcripts Claude Code writes locally.

Can I use the dashboard without the mobile companion? Yes. The TUI is the primary interface. The PWA in cmd/web/ is optional and runs separately via make web.

How is this different from a generic tmux session manager? A session manager creates and switches panes. agent-dashboard understands what's running in each pane — it parses Claude Code's JSONL transcripts to detect state (blocked, waiting, running, done, PR, merged), captures plans and Mermaid diagrams, tracks token usage, and integrates the GitHub PR workflow.

Is this related to Claude Code's official UI? No. agent-dashboard is an unofficial third-party plugin. It builds on top of Claude Code's hooks system and JSONL transcripts but is not affiliated with Anthropic.

Does it work on Windows? Native Windows is unsupported — the project targets macOS and Linux because tmux is required. WSL with tmux installed should work but is untested.