A curated guide to the Herdr ecosystem: tools, workflows, configs, clients, skills, and integrations.
Awesome Herdr 
A curated index of the Herdr ecosystem — the tools people build on top of the terminal-native agent multiplexer.
Herdr is tmux for AI agents. It gives agents and humans persistent workspaces, tabs, and panes; tracks what every agent is doing; survives detach/reattach and remote attach; and exposes a local Unix socket so anything can drive it.
curl -fsSL https://herdr.dev/install.sh | sh
herdrEverything below builds on those primitives — running fleets of agents side by side, wiring editors and MCP clients into panes, switching and restoring sessions, and shaping the terminal around them.
At a glance
| Group | Project | What it gives you | |---|---|---| | Orchestrate | 
Run & orchestrate agents
The headline use: more than one agent working at once, in panes you can watch.
The official, upstream skill file: drop-in instructions that teach an agent already running inside a Herdr pane to use the multiplexer — list workspaces, tabs, and panes, spawn helpers, send input, and wait on output and agent status. The canonical starting point before reaching for any of the third-party skills below.
Walks your live agent tabs and drops a conductor on each — a per-tab technical-PM that reads the session, cross-checks git, surfaces scored impact×effort decisions, then drives the agent already sitting there to carry them out. It conducts rather than codes: one manager per tab (or split beside the agent), each in isolated context, so you stay on the decisions instead of babysitting ten parallel projects. Backend-agnostic across Claude, Codex, Pi, and Hermes, with an optional persistent .herdr-pm/ notebook for the repos you want a standing PM on.
A skill that teaches Claude, Pi, or Codex to spawn a named peer agent, prompt it, and read its output when it goes idle — so an agent can delegate the way you would. A shell wrapper smooths the rough edges, sending Codex's composer a second Enter and falling back to a manual pane split when herdr agent start loses the process before detection.
A personal skill collection whose herdr-pair skill runs Claude and Codex as co-equal peers in one Herdr tab: either can initiate, the other auto-joins, and they exchange structured task / review / question / accepted messages until both sign off. A solid reference for how a real peer-agent protocol — spawn checks, send verification, per-tab session isolation — is wired against the Herdr CLI.
Turns a workspace into a small team: an overseer hands work to implementer, tester, reviewer, and researcher agents, each in its own Herdr workspace and Jujutsu worktree. Every command a worker runs is checked against its role's allowlist — testers can run tests but not write files, pushes need a one-time human-approved token — and all task state persists under .pi/overseer/ so a run survives a restart. For anyone who wants structured, guard-railed multi-agent execution rather than one chat at a time.
Turns Pi-in-Herdr into a gated pipeline: an orchestrator spawns planner, coder, and reviewer agents in their own panes, requires the plan to be approved before any code is written, and runs a review on each phase before it lands as its own commit. Handoffs pass through files under .pi-herdr/ instead of terminal pastes, keeping long plans and diffs out of the input stream. For Pi users who want enforced plan-then-review structure over free-form agent runs.
A Pi extension that gives Pi's background-task and conversation-fork features a home in Herdr: /bg offloads the running task into a visible pane, and /tab forks the conversation into a new tab, both over the socket API. Long builds and parallel workstreams become first-class panes and tabs you can watch, instead of invisible background processes. For Pi users running inside Herdr who want their side-work surfaced, not hidden.
A maintained suite of Pi coding-agent extensions from Herdr's creator, spanning ephemeral overlays, parallel-agent spawning, session recall, and — through the pi-herdr package — direct orchestration of Herdr panes, tabs, and workspaces from a Pi session. Install any package individually with pi install npm:@ogulcancelik/<name>. The reference collection for Pi users who want first-party Herdr integration alongside a rich set of productivity extensions.
A Pi extension that delegates work to visible subagents in a shared Herdr tab named subagents, so you can watch each delegated task directly instead of waiting for a result. Parent Pi sessions get tools to spawn named child agents, steer them mid-run, and read their output when they go idle. Requires herdr integration install pi and Pi running inside a Herdr-managed pane.
Turns the main Claude Code session into a coordinator that plans, dispatches workers to Herdr panes, and monitors their output — without writing code itself. Includes Python monitoring helpers, a YAML state-machine template, and prompt files for implementation, verification, and review roles. For teams who want enforced plan-then-execute discipline in Herdr without building their own multi-agent harness.
Adds natural-language workspace control on top of Herdr's official skill: say "open a worktree for feat-login off main" and the agent translates it into the right herdr CLI commands, including worktree operations the upstream skill omits. Uses progressive disclosure — a lean primary document plus referenced detail files — to stay light on context. For developers who find the raw Herdr CLI syntax tedious to compose in conversation.
Spawns tester, coder, and reviewer agents in dedicated Herdr tabs and orchestrates them through local task cycles and GitHub PR reviews, storing all state in files rather than session context so runs survive restarts. A human-approval gate and run log keep the loop auditable; a budget cap and pause file let you stop it mid-cycle. For developers who want automated PR review cycles that stay observable and interruptible.
Shims Claude Code's experimental agent-teams feature onto Herdr so teammates spawn as native Herdr panes rather than tmux panes, using a translation layer between tmux commands and the Herdr socket API. Requires Herdr 0.6.10 and integrates with the sidebar, metadata, and notification surfaces natively. For Claude Code users who want visible team-mode panes without tmux inside a Herdr workspace.
An official plugin from Herdr's creator that turns a GitHub issue, PR, or discussion into a ready-to-work agent tab — it creates the tab, starts Codex or Claude, renames the session, and sends a structured prompt describing the linked item, all from one bound keypress. It accepts short references like #614 or issue 614 as well as full URLs. For developers who kick off agent sessions straight from a GitHub ticket without assembling the context by hand.
A first-class Herdr plugin built around two accelerators: Projects, declarative TOML templates that spin up an entire workspace — every tab, pane, and startup command — from a fuzzy picker in one keypress; and Quick Actions, a fuzzy launcher for one-off scripts in the current directory. It installs with or without a local Go toolchain and ships prebuilt binaries. For power users who want one-keypress environment setup across many repositories and a fast palette for recurring tasks.
A loop-engineering toolkit that layers three kinds of bounded agent iteration on Herdr: an inner loop that converges a single task through a mechanical gate and an adversarial cross-model judge, a fleet layer that runs batches of tasks in parallel worktrees, and an epic layer that decomposes a large goal and integrates the results into a branch. All state lives on disk so an interrupted run resumes cleanly. For developers who want auditable, convergence-checked agent loops rather than open-ended runs that stall silently.
A Claude Code skill that processes GitHub issues end-to-end inside Herdr: each issue gets its own worktree, a Claude worker implements and pushes a draft PR, then paired Codex and Claude reviewers in split panes run fix-and-review rounds until the diff is clean and auto-merge is armed. A dashboard pane tracks per-issue stage and elapsed time across everything running in parallel. For teams who want a hands-off issue-to-merge pipeline that stays observable in the terminal.
An autonomous coding factory that claims items from Jira or local markdown task files, spins up a Herdr worktree per item, and runs Claude Code through an ordered pipeline of steps (fix → review → PR, or a custom belt), riding each PR through CI and review to merge under a global concurrency cap. Belts are YAML-configured with their source, steps, match predicates, and priority. For teams who want a walk-away pipeline built on Herdr's workspace and worktree primitives.
A headless orchestrator (docs in Japanese) that polls a GitHub Projects board for candidate issues, creates worktrees with gwq, and runs an opencode agent per issue in a Herdr pane, using herdr agent wait to detect completion. Agent state and logs live entirely in Herdr workspaces rather than a separate dashboard. For teams who want a background issue-to-PR pipeline driven from a project board.
Brings the AgentBox sandbox into a Herdr session: a live overlay on prefix+a, a one-key shortcut to start a sandboxed VM in the current project, and an agentbox:// Ctrl-click link handler. The plugin installs from the Herdr marketplace and wires its keybindings through a generated shim so the manifest stays static. For Herdr users who run agents in AgentBox sandboxes and want to reach them without leaving the terminal.
A Pi package that exposes Herdr's agent, pane, and session control surface as Pi slash commands and LLM tools — start, send, read, focus, and stop agents without leaving a Pi conversation. It's deliberately generic runtime plumbing that product-specific loop extensions can build on rather than reimplement. For Pi users who want to drive Herdr orchestration from inside Pi.
A Pi extension that hooks Pi's async subagent event bus, tracking each run's id, task, status, and session file, and exposes a TUI dashboard where you can browse live subagents, open their session in a new tab, focus an existing pane, or spawn one directly into a pane. Unlike the shared-pane model of aldrickdev/herdr_subagents, it watches without changing how subagents run. For Pi users with many concurrent subagents who want one navigable view of the whole fleet.
A Pi extension that spawns each subagent into its own Herdr pane or tab, manages the session files, and notifies the parent session on completion so the orchestrator picks up results without polling. Child panes stay open after reporting back, leaving room for follow-up work, and each subagent can run on a different model. A dedicated-pane alternative to the shared-pane aldrickdev/herdr_subagents.
An OpenCode plugin that watches for subagent session creation and automatically splits a new Herdr pane running opencode attach for each one, tiling them in a row-based grid to keep the layout readable. Panes can close automatically on idle, deletion, or error. For OpenCode users who want live visibility into every active subagent without splitting panes by hand.
An installable skill that teaches Claude Code or Hermes to orchestrate a fleet of coding agents through Herdr — discovery, spawn, dispatch, fan-out/converge, approval unblocking, and spec-driven loops where tasks.md markers map directly to parallel workers. An onboarding TUI sets up the whole factory in one pass across Claude, Hermes, or Cursor. Broader than a coordinator-only skill: it adds steered and meta-orchestration tiers and integrates spec-kit.
A Hermes and Claude Code skill that spawns an interactive Claude Code (Fable 5) worker inside Herdr, hands it the prompt through a file, and reads the complete answer back — sidestepping TUI scraping and output truncation. The worker's session UUID is surfaced so a follow-up can resume the same conversation. For users whose primary agent is a lighter tier who want to route deep-reasoning tasks to a more capable model without switching sessions.
A refinement of the default Herdr coordination patterns that swaps vague wait agent-status done polling for output-marker waits and structured messages carrying sender pane, reply-to metadata, and task kind — so a receiver can answer directly without the sender blocking. Ships a dependency-free herdr-msg Bash helper and a CLI reference behind progressive disclosure. For multi-agent workflows where blocking waits create needless stalls between sibling agents.
A Claude Code plugin bundling five orchestration skills — a full-lifecycle "boss-mode" conductor, a one-shot Herdr fan-out, the Herdr CLI reference, an end-of-session durability auditor, and a docs scaffolder — alongside hooks that guard worktree hygiene, route prompt events, and report session state. Boss-mode owns a backlog of parallel initiatives end-to-end: spawn, build, QA gate, merge, deploy, verify-live, teardown. For Claude Code users who want a pre-wired multi-skill setup rather than assembling the pieces.
A live timeline panel — dockable at the bottom or right — that aggregates agent task events across every Herdr workspace, showing working and blocked states, session IDs, and a deduplicated 7-day history, with individually togglable columns. It follows agents across workspace switches without a manual refresh. For anyone running several agents at once who wants a single panel to see which are active, blocked, or idle.
A Bubble Tea TUI for prompts you want to run later: jot multi-line entries with optional titles, fuzzy-filter them, then press Enter to paste one into a running Claude Code pane or spin up a fresh claude tab in the current workspace. Per-repo backlogs live under <repo>/.herdr-todo/ so they travel with the project, alongside a global backlog visible everywhere. For engineers who accumulate follow-up prompts mid-session and want to queue them without interrupting the agent.
A bridge that turns an iPhone Action Button into a remote for your own Herdr agents: a bundled iOS Shortcut with Whisper dictation sends a voice-transcribed task over Tailscale through a personal Telegram bot to a Claude Agent SDK operator, which starts or targets the right Herdr lane and replies in Telegram. A self-bootstrapping skill handles the one-time setup. For developers who want to assign work to their Herdr-backed agents from anywhere without opening a laptop.
Self-hosted mission control for interactive coding agents: spawn, watch, and steer many real Claude Code or Codex sessions in parallel from a browser or phone, each running in its own git-worktree pane that Herdr multiplexes. Plan-review, PR-critic, and merge-train gates layer engineering discipline on top of the parallelism. For teams running several agents at once who want a supervised fleet rather than a wall of terminals.
Launches a coding agent fully sandboxed in a libkrun/KVM microVM straight from a Herdr pane, while Herdr keeps its usual multiplexing and detach/reattach UX around the isolated process. A baked agent image and a shell dispatch pipeline handle the VM lifecycle. For Linux users who want hardware-level isolation per agent without giving up Herdr's pane control.
A Go control-plane daemon that drives the issue-to-PR loop from a declarative YAML state-graph, using Herdr as the execution backend — it spawns an implementer agent in a worktree, then a reviewer, then polls GitHub's merge gate before squash-merging (dry-run by default). Workflow configs are JSON-Schema-validated and checked with a validate subcommand. For teams who want a deterministic, auditable pipeline around agent-driven PRs.
Provisions an isolated worktree per branch or GitHub issue — its own ports, .env, and setup — and boots a coding agent into a Herdr pane for each, then tracks status, lands finished branches, and tears them down. It shells out to the Herdr CLI and only runs inside an active Herdr session. For developers running several agents or issues in parallel who are tired of services fighting over port 3000.
Two companion scripts that spin up a project-scoped Herdr workspace with an N-agent grid (default 2×2, one Pi agent per pane) keyed to the current directory, and relaunch any dead agents when you return. For anyone who wants a repeatable one-command "start a grid of agents for this repo" setup.
Connect over socket & MCP
Wiring outside tools — Python code, MCP clients — into Herdr.
A zero-dependency Python client for Herdr's Unix socket, so you can script the multiplexer instead of pressing keys. It handles the tedious parts — socket discovery, request envelopes, typed errors, event subscriptions, pane reads and waits — with ten convenience helpers and a raw request() escape hatch validated against the full method surface. The fastest way to build your own automation.
An MCP server that exposes Herdr to any MCP-speaking client — Claude Desktop, Cursor, Claude Code — as 21 tools for discovering, launching, reading, and writing panes. A recipe engine chains those calls into reusable flows with {{ step.result.path }} variable passing, reachable over both MCP and an HTTP bridge, with a bundled React playground for trying them out by hand.
Another MCP server, this one tuned for agents coordinating agents: it hands any MCP-capable client tools to read another agent's pane, relay a message, hand off a task, spawn an agent, and wait on a result. It turns manual copy-paste between panes into one-sentence orchestration, and herdr-mesh install auto-registers itself with whichever agents (Claude Code, Codex, opencode) you have installed.
Publishes Codex subscription quota — remaining 5-hour and weekly usage — into the Herdr agents sidebar as compact labels refreshed every 30 seconds, and opens a live usage dashboard in any shell pane. It reads Codex's own JSONL session logs without calling an API or consuming tokens, and installs as a systemd service plus two terminal commands. For Codex users running on Linux who want quota awareness without leaving the Herdr workspace.
Official reference plugins from Herdr's creator demonstrating four patterns: Telegram notification, development layout, GitHub link preview, and Rust release tracking — each a standalone herdr-plugin.toml package. They are provided as-is for adaptation rather than direct dependency. The canonical starting point for developers building their own Herdr plugins before reaching for the full API docs.
Maps each live Herdr pane to a Telegram forum topic so your AI agent activity streams into a chat thread readable on any device. Accepts bot commands to relay input back into panes and consumes structured turn data from Herdr when available. For developers who want ambient monitoring of their Herdr sessions through Telegram without keeping a terminal window visible.
A single-binary MCP server that exposes Herdr's socket API as MCP tools with no state of its own — each call opens a fresh socket, sends one request, and returns. It covers 75 methods across workspace, tab, pane, agent, layout, and plugin surfaces, with named profiles (coordinator, client, observer) to scope the tool set per agent role and HERDRMCPDENY globs for finer removal; an unknown profile falls back to read-only. A leaner, role-aware alternative to the recipe-engine herdr-mcp for wiring agents straight to Herdr.
A Go module that speaks Herdr's newline-delimited JSON socket protocol directly, generated against the 0.7.0 surface, with typed helpers for common calls (current pane, pane read, agent-status subscribe) and a Call/CallRaw escape hatch for the rest. Socket resolution follows the standard env-var and default-path chain. The Go counterpart to the existing Python socket client, for developers writing plugins, CLIs, or automation in Go.
A plugin that listens to Herdr lifecycle events and emits an NDJSON stream of pane focus intervals, detected agent and model metadata, and local session-trace summaries to a file, an HTTP webhook, or any command's stdin. Raw transcript text is redacted by default, with opt-in controls for trusted local sinks. For developers building menu-bar apps, time-trackers, or dashboards who want structured Herdr activity without writing their own socket client.
A user-level MiMo Code plugin that reports idle / working / blocked / done state to Herdr's sidebar via pane.report_agent, using an aggregate state machine across all MiMo subagent sessions to avoid flicker during multi-session runs. A detached watchdog releases the pane label on any exit — including kill -9 and hard crashes — and a subprocess guard suppresses the plugin when MiMo runs as a tool call inside another pane. For MiMo Code users who want their agent's status visible in Herdr without waiting for native support.
A plugin that pushes a structured notification to any ntfy topic when an agent blocks or finishes, so alerts reach your phone, tablet, or desktop regardless of which machine is running the agents. Each message carries the workspace, tab, and pane ID so you know exactly where to return, and a local ntfy server is auto-detected for near-instant delivery before falling back over the network. For developers who want cross-device agent alerts through the open ntfy ecosystem.
A plugin that emails you a one-line summary when an agent blocks, finishes, or fails — and lets you reply to that email to drive the agent forward, turning your phone's mail app into a remote control for sessions running over SSH. A reply-watcher polls the tinysend mailbox and feeds your answer into the right pane to resume the blocked agent. For developers who manage agents remotely and want a low-friction approval loop from any device with email.
A plugin that fires a clickable macOS notification when an agent goes blocked or done — but only when you're away from that pane, so you're never alerted about a pane you're already watching. Clicking the toast brings the terminal forward and focuses the exact pane that needs you, via the alerter utility. For macOS users who leave agents running unattended and want to be pulled back to precisely the right pane.
A plugin that ships its own branded HerdrNotify.app so agent alerts carry the Herdr logo in macOS Notification Center — not a generic terminal icon — with templated messages and a click action that jumps to the blocked or finished pane. The bundled app re-registers with Launch Services on a six-hour TTL, recovering the icon automatically after reboots or OS updates without a cron job. For macOS users who want polished, on-brand desktop notifications from their agent fleet.
A zero-dependency plugin — only curl and system Python or jq — that forwards agent status events to the herdr-remote relay whenever an agent blocks or changes state, enabling mobile monitoring and one-tap approval from the herdr-remote app, menu bar, or Telegram bot. Configuration is a single HERDR_RELAY variable, and a built-in test action verifies the pipeline without waiting for a real event. The Herdr-side companion to herdr-remote (listed below).
A Chrome extension plus local daemon that watches which browser URL was focused while you spoke to the Hex voice transcriber, then routes the resulting transcript to the matching Herdr workspace and pane — no manual focus switch. Routing rules map URL patterns to workspace/tab/pane triples and are configured in the extension popup. For developers who dictate to several Claude Code panes and want each spoken instruction to land in the right agent automatically.
A Letta Code modification that reports accurate idle / working / blocked state to Herdr's socket when the agent runs inside a Herdr pane, so a Letta session shows up in the status sidebar like any native agent. For Letta Code users who want the same pane-status visibility Herdr already gives other agent CLIs.
Wires TraeX's lifecycle hooks to Herdr's pane.reportagent / pane.releaseagent socket calls, so a TraeX pane shows correct idle / working / blocked status in the sidebar even though Herdr ships no built-in detector for it. For anyone running the TraeX CLI agent inside Herdr panes.
Keeps a Herdr pane flagged as working — with a ⏳ label — while background CI, review, or
README truncated. View on GitHub
