Anyone who works seriously with Claude Code or Codex rarely has just one session open. One agent builds a feature, a second writes tests, a third cleans up a migration, and somewhere a fourth has been waiting twenty minutes for an approval. The question is no longer which model writes the code, but how you keep track: which session needs me, which branch is it working on, and what does all of this cost?
Since 2025, a distinct class of tools has emerged for this problem, and in 2026 it became widely established. Two representatives with very different approaches are herdr and Orca. This article compares both factually, places other well-known tools in context and describes my own setup. At the end it turns to the question none of these tools answers: how does a company see how many tokens and how much agent time its developers use, and for which project?
In short: herdr and Orca are free, open-source tools for running several coding agents such as Claude Code and Codex in parallel. herdr is a terminal runtime with a native app that keeps sessions alive and leaves the agents unchanged. Orca is a development environment that starts every task in its own Git worktree, brings review and pull requests, and skips the agents’ permission prompts by default.
Disclosure: I work with herdr every day. I describe Orca based on its official documentation, repository and changelog. All information reflects the state of 22 November 2026.
Harness, orchestrator, gateway: three different layers
Before looking at tools, it helps to clarify terms, because the three layers are often confused.
The harness is the program an agent works in: Claude Code, Codex, OpenCode or Pi. It runs the loop of model call, tool call and result. We described what a harness does in AI harness explained, and you will find an overview of common harnesses under AI harness.
The orchestrator starts and holds several harness sessions, shows their state and organizes the working directories. This is where herdr and Orca play. Neither replaces Claude Code or Codex; both launch their CLIs.
The gateway sits between the harness and the model provider. It controls who may use which model, checks budgets and attributes costs. That is the layer where we work with the innFactory AI Gateway.
Keeping these layers apart avoids a typical mistake: expecting cost control from an orchestrator, or a better terminal interface from a gateway.
What is herdr? Terminals that never let go
herdr is an open-source runtime for coding agents that runs agents in long-lived terminals, shows their state and does not modify them. herdr comes from Herdr, Inc., a startup from the Y Combinator program that closed a USD 6 million seed round led by Bessemer in September 2026. Today the product consists of three programs; according to the FAQ, the app and the TUI explicitly do not share agents:
- Herdr (app): a native desktop app rendered with GPUI, for macOS (Apple Silicon only), Linux and Windows, plus apps for iPhone, iPad and Android. The current version is 0.1.16.
- herdr-core (
hc): the terminal runtime beneath the app. The app is just one client; the API can be printed as an OpenRPC description withhc schema. - Herdr TUI: the original herdr as a terminal application with its own server, currently v0.9.5 and, with more than 52,000 stars, the maker’s best-known repository.
The core idea is persistence. Agents run in real terminals inside herdr-core. Quit, update or crash the app, and the terminals keep running. A reboot of the machine does end the processes; the TUI then restores the layout and supported agent sessions, while the app does not restart agents on its own.
The app detects agent state through an open terminal protocol (OSC 7501, “Program Status Protocol”), according to the maker without screen scraping and without a list of known agents. The TUI instead reads state from the screen via detection manifests or takes status reports from integrations and agents, and knows more than 20 agents, including Claude Code, Codex, GitHub Copilot CLI, Cursor Agent, OpenCode, Pi, Gemini CLI and Amp. Agents can start helpers themselves via hc run --bg, hc send and hc wait and collect their results.
herdr supports Git worktrees, but not as its organizing principle: in the app, a bundled mod adds creating, opening and removing worktrees; the TUI offers methods for this in its socket API. Further machines join via SSH (hc machine add, hc run --on). According to the maker, this needs no account, and terminals never pass through Herdr’s servers.
Its most striking trait is extensibility. The app’s entire interface lives as files under ~/.config/herdr-app: layouts, key bindings, themes and mods. Mods are JavaScript that describes the interface, which is rendered natively. They can run programs and fetch URLs. According to the homepage, the gallery holds 576 mods and 312 layouts.
All three programs are licensed under Apache-2.0 and free. Pricing, team features or an enterprise offer are not documented; a service called “Herdr Cloud” was announced in September 2026 as a connection service that explicitly does not host the agents itself. The desktop app sends an anonymous usage report once a day with counts, agent and layout names, without terminal contents or prompts. It is on by default, can be switched off and is kept for three months.
What is Orca? A development environment for agents
Orca is an open-source development environment for coding agents that starts every task in its own Git worktree and covers the way to the pull request in one interface. Orca comes from Stably AI (Lovecast Inc.), also a Y Combinator company from San Francisco. Orca calls itself an “Agent Development Environment” and is an Electron-based desktop app for macOS (Apple Silicon and Intel), Windows and Linux, complemented by a mobile companion app. The stablyai/orca repository is MIT-licensed, has more than 132,000 stars and ships several releases per week; the current version is v1.4.269 from 21 November 2026.
The organizing principle is “worktree-first”: every task gets its own Git worktree and branch. Files that Git ignores, such as local environment variables, can be carried into new worktrees via configuration. Around the agents, Orca builds a complete working environment: a GPU-rendered terminal with splits, an editor, an embedded Chromium browser with a design mode, and a diff review in which you collect line comments and hand them back to the agent in one go. Commit, push and pull requests for GitHub, GitLab and Bitbucket Cloud, including checks, are built in, with auto-merge and stacked PRs on GitHub, as are integrations with Linear and Jira.
Orca knows more than 40 preconfigured agents. Integrated more deeply, with usage display, hooks and, depending on the agent, account switching, are Claude Code, Codex, Cursor and Muse Code. Orca is explicitly built for subscriptions: “Plug in the subscriptions you already have.” Claude Code uses the existing login, several Claude and Codex accounts can be stored and switched with a click, and usage and rate limits appear in the interface.
For remote machines, Orca offers four modes: local, via SSH (agents and Git run on the host, and Orca installs its own server there), as a Remote Orca Server in beta, and experimentally as a cloud VM per workspace with your own provider. Through “Orca Bots”, tasks can be accepted from Slack, Linear, GitHub or via webhook; these connectors run through Orca Cloud.
One point deserves attention: according to its documentation, Orca pre-fills the agents’ bypass flags for new launches, that is --dangerously-skip-permissions for Claude Code and --dangerously-bypass-approvals-and-sandbox for Codex. The rationale is that worktrees are disposable. A worktree, however, only isolates files in the repository, not processes, network access or credentials on the machine. You can switch to “Manual” in the settings. For companies, this should be the first setting they decide on deliberately.
On data flow, Orca states that prompts and code go directly to the configured providers. Product telemetry is on by default for new installations, can be switched off and goes pseudonymized to PostHog in the US. The mobile connection via the relay requires a sign-in and, according to the privacy policy, uses infrastructure in the US and Hong Kong. Enterprise terms are available on request; documented there is “SOC 2 Readiness”, not a certification.
herdr vs. Orca: the direct comparison
| Criterion | herdr | Orca |
|---|---|---|
| Approach | Terminal runtime with a native app, focused on persistence and customization | Development environment for agents, focused on the way to the merge |
| Interface | Native app (GPUI), TUI, mobile apps | Electron app with terminal, editor, browser, diff and chat |
| Worktrees | Supported (mod in the app, API in the TUI) | Organizing principle: one task, one worktree |
| Agents | Any terminal CLI; the TUI detects more than 20 agents | Any CLI; more than 40 preconfigured |
| Subscription use | CLIs run unchanged with their own login | Existing logins, account switching, usage display |
| Permissions | The respective agent’s settings | Bypass flags by default, switchable to Manual |
| Review and PR | Panels for files and changes, no PR workflow documented | Diff review, commit, PRs, Linear and Jira |
| Remote | SSH directly, without Herdr servers in between | SSH, remote server (beta), cloud VM (experimental) |
| Platforms | App: macOS (Apple Silicon), Linux, Windows; TUI: macOS, Linux, Windows | macOS (Apple Silicon and Intel), Windows, Linux |
| License and price | Apache-2.0, free | MIT, free, enterprise on request |
| Extensibility | JavaScript mods and layouts, plugins in the TUI, OpenRPC API | Plugins, CLI, automations, bots |
| Team | Not documented | Bots with an allow list, shared remote runtime |
| Data flow | Terminals never pass through Herdr servers, anonymous telemetry can be switched off | Model traffic direct, telemetry on by default and can be switched off, cloud for bots and mobile relay |
You will find profiles of both tools with all supported agents in our AI harness overview under herdr and Orca.
Alternatives to herdr and Orca: Conductor, Paradev, cmux, Claude Squad
herdr and Orca are not alone. Four more tools come up regularly:
- Conductor is a desktop app for macOS, complemented by Conductor Cloud, an iOS app and an API. Every workspace runs in its own worktree; it supports Claude Code, Codex, Cursor and OpenCode, and Claude Pro/Max and “Sign in with ChatGPT” work. Locally, agents run without a sandbox per its FAQ; cloud workspaces run in isolated sandboxes. In early November 2026, Meta acquired Conductor; the initially announced shutdown was reversed on 13 November, and releases have been shipping regularly again since.
- Paradev, called Superset until October 2026, is a desktop app with one worktree, branch and terminal per task. Like Orca, it launches Claude Code and Codex with bypass flags by default per its docs. Paradev is MIT-licensed, with a free individual plan and paid plans from USD 16 per user and month. macOS is the main platform, Linux is experimental.
- cmux by Manaflow is a native macOS terminal app with vertical tabs, notifications and a browser pane. Worktrees are not built in. The app is licensed under GPL-3.0, the server components under BUSL-1.1; besides a free plan there are Pro, Max, Team and Enterprise plans, which mainly add cloud VMs.
- Claude Squad is a lean terminal application written in Go on top of tmux and worktrees, AGPL-3.0, free.
The providers themselves are catching up as well. Claude Code Desktop runs parallel sessions in a sidebar, optionally with a worktree per session, locally, via SSH or in Anthropic’s cloud. Codex is now part of the ChatGPT desktop app and also uses worktrees for parallel chats there. If you only use one provider, you do not necessarily need a separate orchestrator. herdr and Orca make the difference where several harnesses run side by side.
herdr or Orca: which tool fits whom?
herdr fits developers who are at home in the terminal, work a lot on remote machines and want to shape their interface themselves. Its strength lies in robust, long-lived sessions and in the fact that herdr does not modify the agents: permissions, hooks and configuration remain those of the respective CLI. If you expect a complete review and PR workflow in one interface, you will not find it ready-made in herdr.
Orca fits teams that see agent work as a pipeline: task, worktree, agent, review, PR. This discipline is built in and saves manual work. Before a broad rollout, you should settle three points: the default permissions, the cloud components for bots and mobile relay in your records of processing, and the telemetry setting.
Regardless of the tool: orchestrators make parallelism cheap. Three agents instead of one are a keystroke away. Costs still grow linearly, and with subscriptions nobody in the company sees them. More on that below.
My setup: herdr, Claude Code and Codex on an Azure VM
I do not code on my laptop. My development environment is a VM in Microsoft Azure that I reach via VPN. Code, agents, builds and local servers run there. When an agent starts a development server, such as a Hugo preview or a web interface, I open it through the VPN tunnel directly in my Chrome, as if it ran locally. The benefit: agents can run for a long time without the laptop having to stay awake, and all sessions live in one place.
On top of that runs herdr as the runtime for all agent sessions. In the team, we share our own herdr mods in a common repository. I use three of them every day: a project layout that shows sessions in folders instead of tabs and marks at a glance which agent is working and which one needs me; a resume mod that brings back running Claude Code and Codex sessions after a restart with claude --resume or codex resume; and a usage view that shows tokens and estimated costs of Claude Code per day, model and project. Such mods are quick to build in herdr, because a coding agent can write them itself using the bundled manual.
Inside herdr run Claude Code and Codex. I use Claude Code both through a subscription and through our AI Gateway. Codex has its own profile that calls the Responses API through the gateway. The gateway’s companion aigw also runs on the VM as a background service. For Claude Code and Codex it records calls, exact tokens, agent time and the tool version, including sessions that run on the subscription and never touch the gateway. It does not upload prompts.
A second brain inside the coding agent
Next to the coding agents runs Saibot, my personal assistant. Saibot is my second brain: notes, decisions, emails, documents and repositories are searchable there, and agents look things up there before they claim anything about projects or past decisions. Saibot only reads and sends nothing; it proposes actions in a queue that I approve. That goes beyond this comparison, but one detail is interesting for developers: the connection to Claude Code through hooks.
Claude Code supports hooks for events in a session’s lifecycle. I use four of them:
{
"hooks": {
"SessionStart": [{ "matcher": "startup|resume|clear|compact",
"hooks": [{ "type": "command", "command": "$HOME/.claude/klaus/sessionstart.sh" }] }],
"Stop": [{ "hooks": [{ "type": "command", "command": "$HOME/.claude/klaus/stop.sh" }] }],
"SessionEnd": [{ "hooks": [{ "type": "command", "command": "$HOME/.claude/klaus/stop.sh" }] }],
"PreCompact": [{ "hooks": [{ "type": "command", "command": "$HOME/.claude/klaus/stop.sh" }] }]
}
}When a session starts, the hook asks Saibot for what has been learned about this project. It sends the project name and the list of files currently changed or new in the Git working tree. That way, insights about exactly those files come first, not just the newest ones. The answer lands in the session as additional context: where the last session left off, which pitfalls and decisions came up, each with a key through which the agent loads details on demand.
On stop, at the end of a session and before every compaction of the context, the second hook cuts off the new part of the transcript since the last delivery and sends it to Saibot. The hook itself evaluates nothing. Saibot condenses each delivery in a background job with one inexpensive model call per delivery, not per tool call. The hook before compaction matters for long sessions that never end cleanly: without it, everything would arrive at once at the very end.
Two design decisions have proven themselves. First, both hooks work “fail-open”: if Saibot is unreachable, the session still starts, and the error only lands in a log. A hook that blocks sessions or starts them with error messages gets switched off, and then nobody learns anything anymore. Second, the start hook has a hard limit of two seconds. The result in daily work: a new Claude Code session knows what the previous one learned about this repository, for example which build step tends to fail or why a decision was made the way it was.
How do companies measure the tokens and agent time of coding agents?
For an individual developer, it is enough if the tool shows how much of their own subscription is left. Orca can do that, and herdr can via mods. A company asks different questions: how many tokens do our developers use in total? Which project, which customer, which cost center causes them? How much work do agents actually do, and what would the subscription usage be worth at list price? None of the orchestrators answers this across a team, and it is not their job either.
The innFactory AI Gateway closes this gap with two operating modes that can be combined.
In proxy mode, Claude Code, Codex, OpenCode and other tools talk to the gateway in their native dialect: OpenAI, Anthropic or Gemini. The gateway checks budgets, permissions and guardrails, picks the provider and attributes every token to a user, a repository and a cost center. Ready-made setups exist for Claude Code, Codex, OpenCode, Cowork and VS Code; aigw run launches a tool with a short-lived token without changing its own configuration.
In companion mode, the aigw service on the developer’s machine measures what happens beside the gateway. It reads the local session logs of Claude Code, Codex, OpenCode, Pi, GitHub Copilot CLI and Gemini CLI and reports agent time, LLM time and tokens, including subscriptions such as Claude Max or ChatGPT. Subscription usage is valued at list price but never billed or counted against budgets. Rules by repository, organization or host attribute agent time to projects and cost centers. Prompts, absolute paths and credentials never leave the machine. The companion installs with one line or fleet-wide via MDM.
Both modes work regardless of whether a session runs in herdr, in Orca or in a plain terminal. The gateway sits one layer below. For organizations this means: you can leave the choice of orchestrator to your developers and still see how many agent hours flow into which project. The license starts at 199 EUR per month, and the gateway runs in your own cloud.
Conclusion
herdr and Orca solve the same problem with different philosophies. herdr relies on long-lived terminals, minimal intervention in the agents and an interface you shape yourself. Orca relies on a guided environment from task to pull request, with worktrees as the organizing principle and subscriptions as the norm. Both are open source and free, and both move fast. Before rolling either out to a team, check the default permissions, the telemetry and the cloud components.
The real question for companies lies one layer below: who sees what the agents cost and which project they work for? If you want to capture your developers’ tokens and agent time and attribute them to projects, including subscription usage, we are happy to show you the AI Gateway on a live system. Request a demo.
