↓ Skip to main content
9 – 17 UHR +49 8031 3508270 LUITPOLDSTR. 9, 83022 ROSENHEIM
DE / EN
Orchestrator Superset Inc. USA Desktop app, CLI, mobile, cloud

Paradev

Paradev (formerly Superset): MIT-licensed desktop app that runs Claude Code, Codex and other coding agents in parallel in separate Git worktrees.

Licence MIT
First release December 2025
Interfaces Desktop app (Electron), CLI (superset or paradev), MCP server and TypeScript SDK, iPhone and iPad app, Cloud workspaces
Platforms macOS, Linux (experimental)

The seven building blocks in Paradev

How Paradev implements the building blocks of an AI harness

  1. 1

    Agent loop

    Paradev does not run an agent loop of its own; it launches the installed agent CLIs in the terminal. Claude Code, Codex, OpenCode and Pi can also be shown as a chat via the Agent Client Protocol, with the same login and the same MCP servers.

  2. 2

    Tools and execution environment

    Tools come from the respective agent; Paradev adds a CLI, an MCP server, a TypeScript SDK, skills, a plugin marketplace, an in-app browser with design mode and computer control for native desktop apps.

  3. 3

    Context management

    Every task runs in its own workspace with a Git worktree, branch, terminals and ports under `~/.paradev/worktrees/`; setup and teardown scripts from `.paradev/config.json` prepare new workspaces.

  4. 4

    State and memory

    A background daemon keeps terminals, processes and scrollback alive across app restarts and updates; sessions of supported agents can be resumed, forked or handed to another agent with their transcript.

  5. 5

    Permissions

    According to its docs, Paradev starts Claude Code with `--dangerously-skip-permissions` and Codex with `--dangerously-bypass-approvals-and-sandbox`; to keep the approval prompts, remove the flags under Agents in the settings. Per the docs, worktrees do not sandbox processes.

  6. 6

    Verification and feedback

    A diff viewer with line comments sent to the agent, pull requests with checks, review threads, stacked PRs and merge queue, and the in-app browser for previews support the review before the merge.

  7. 7

    Orchestration

    Paradev's core contribution: many agents in parallel, one worktree per task. On top come the `superset:orchestrate` skill, with which Claude Code or Codex starts workers in their own workspaces, and automations that start agents on a schedule or on events from GitHub, Slack, Linear or Sentry.

Models and model access

Which models the harness uses and how they are obtained

Models and model access
Models Model-agnostic: Paradev starts the agents; the models are chosen there
Access routes Model-agnostic (depends on the coding agent started)
Free choice of model Yes
Custom endpoint / LLM gateway Yes
According to its docs, Paradev does not route model calls through its own servers; the agents use your subscriptions and API keys, and the provider bills you directly. For Claude Code, Paradev documents environment variables for Vercel AI Gateway, OpenRouter, Amazon Bedrock, Google Vertex AI, Microsoft Foundry and Anthropic-compatible proxies such as LiteLLM.
Features
MCP servers Sub-agents Hooks Skills Plugins / extensions Headless / CI SDK IDE integration Resume sessions

Enterprise use

Data processing, controls and pricing model according to the vendor

Data processing and location

According to the docs, local workspaces stay on the machine, and model calls go directly to the providers. Remote access via the Paradev relay is off by default. Cloud workspaces run on Vercel machines; all listed subprocessors are based in the US. Per the privacy policy, Paradev stores messages, tasks and attachments you send to agents through its apps.

Permissions

Claude Code and Codex start with bypass flags by default, meaning without approval prompts and without the Codex sandbox. A worktree only isolates files, not processes, network access or credentials. Paradev also installs hooks and wrappers for status display into the agent configurations. For company use, the flags should be removed deliberately.

Telemetry

The desktop app sends usage data to PostHog, linked to account ID, name and email per the privacy policy; errors are processed by Sentry. Paradev does not document a switch, and in the public source code, telemetry is hard-enabled (as of 22 November 2026). Only the SDK's telemetry can be turned off with `SUPERSET_TELEMETRY=0`.

Identity and team

Pro, Team and Ultra allow unlimited members in an organization. Enterprise adds SAML SSO, SCIM, IP restrictions and audit logs. Organization plugins can be distributed by role from a GitHub repository; according to the maker, a SOC 2 Type II report is available.

Pricing

Free costs nothing and covers one user with local workspaces, the desktop app and the CLI. Pro costs USD 20 per user and month (USD 16 billed yearly), Team USD 60 (USD 54), Ultra USD 200; cloud hours are included from Team upwards. Enterprise is priced individually.

Strengths
  • One task, one worktree with its own branch, terminals and ports
  • 22 built-in agents, plus any other CLI as a custom agent
  • MIT-licensed since 4 November 2026, can run on your own Linux server
  • Terminals survive app restarts and updates thanks to a background daemon
  • Usage page with subscription quotas, account switching for Claude Code and Codex and token costs from local session logs
  • Automations, CLI, MCP server and SDK for agent-driven workflows
Limitations
  • Bypass flags for Claude Code and Codex set by default
  • Only macOS fully supported, Linux experimental, Windows without a date
  • Telemetry to PostHog without a documented way to turn it off
  • Remote access, mobile, automations and cloud only in paid plans, services based in the US
  • Cloud workspaces for GitHub repositories only
  • Daily releases: interface and docs change quickly

Use cases

Typical tasks for this harness

Let several coding agents work in parallel on separate tasks, each in its own worktree
Give the same task to different agents and keep the best result
Split large refactorings into independent slices and let a coordinating agent distribute them
Start agents on a schedule or on events from GitHub, Slack, Linear or Sentry
Run agents on your own Linux server and reach them from the desktop app or iPhone

Paradev is an open-source desktop app by Superset Inc. that runs AI coding agents such as Claude Code, Codex or OpenCode in parallel and gives every task its own Git worktree with a branch and terminal. Until October 2026, it was called Superset. As an AI harness, Paradev mainly covers orchestration; model access and tools stay with the agent. As of 22 November 2026.

What is Paradev?

Paradev is developed by Superset Inc.; according to the maker, its three founders were previously CTOs at Y Combinator companies. The first release came out in December 2025. On 26 October 2026, Superset renamed itself Paradev and announced a USD 11.5 million seed round led by Union Square Ventures (announcement). Since 4 November 2026, the code has been MIT-licensed; before that, it was under the Elastic License 2.0. The current version is desktop-v2.0.11 from 20 November 2026; the ParadevAI/Paradev repository has around 16,500 stars (as of 22 November 2026). The app is based on Electron and runs on macOS with Apple Silicon or Intel; a Linux AppImage is experimental, and Windows is planned without a date. The CLI is still called superset, with paradev working as an alias.

Paradev as an AI harness

Paradev organizes work in projects and workspaces: a project is a Git repository, and every workspace is a worktree with its own branch, terminals and ports. According to the FAQ, a worktree is not a sandbox: agents share processes, the network and files outside the worktree. Around the agents, Paradev provides a diff viewer, pull requests with checks, stacked PRs and merge queue, an editor, an in-app browser with design mode and a task list. A background daemon keeps terminals alive across app restarts.

For automation, Paradev offers a CLI, an MCP server, a TypeScript SDK in early alpha and skills with which an agent controls the app itself. The superset:orchestrate skill turns Claude Code or Codex into a coordinator of further agents. Automations start agents on a schedule or on events from GitHub, Slack, Linear, Sentry, Notion, Microsoft Teams or via webhook. Agents run locally, on your own Linux server (“Bring Your Own Cloud”) or in Paradev’s cloud workspaces.

Models and model access

Paradev is model-agnostic: it launches the installed agent CLIs, which determine model, provider and access route themselves, and according to its agent overview, it does not route model calls through its own servers. Subscriptions are explicitly supported: several accounts can be stored for Claude Code and Codex, and the usage page shows subscription quotas and estimates token costs at API list prices from the local session logs. For Claude Code, Paradev documents environment variables for Vercel AI Gateway, OpenRouter, Amazon Bedrock, Google Vertex AI, Microsoft Foundry and Anthropic-compatible proxies. This also lets you route model access through innFactory’s AI Gateway. Per the FAQ, branch names are generated by a small model using your stored credentials.

Company use: data protection and governance

Companies should settle three points before a broad rollout. First, permissions: according to the docs, Paradev starts Claude Code with --dangerously-skip-permissions and Codex with --dangerously-bypass-approvals-and-sandbox. A worktree only isolates files, not processes, network access or credentials, so you should remove the flags deliberately under Agents in the settings. Second, data flows: remote access via the relay is off by default, whereas automations, the Slack integration, Pages and cloud workspaces run through Paradev services. All subprocessors are based in the US, and Vercel provides the cloud machines. Third, telemetry: usage data goes to PostHog, linked to account ID, name and email per the privacy policy; in the public source code, it is hard-enabled. Paradev also installs hooks into the agents’ configurations. For Enterprise, the maker lists SAML SSO, SCIM, audit logs and a SOC 2 Type II report.

Which agents does Paradev support?

According to the documentation, 22 agents are built in, with varying integration depth. Status, resume after restart and usage history are provided by, among others, Claude Code, Codex, OpenCode, Grok CLI, Cursor Agent CLI, GitHub Copilot CLI, Devin CLI, Muse Code and Pi. Gemini CLI, Amp, Droid, Kimi Code and Mastra Code report status and can be resumed. Kiro CLI, Hermes Agent, fx and Polygraph run without status reporting. Account switching and the subagent view are available only for Claude Code and Codex. Any further CLI can be added as a custom agent.

How it fits: Paradev, Conductor, Orca and herdr

Paradev, Orca and Conductor follow the same principle: one task, one worktree, review through to the pull request. Like Orca, Paradev is MIT-licensed; unlike Orca, it requires a paid plan for remote access, mobile, automations and cloud. Conductor supports only four agents and is proprietary, while herdr relies on long-lived terminals without intervening in the agents. You will find a detailed comparison in the article herdr vs. Orca. Paradev’s usage page shows costs per machine; developers’ costs and agent time across teams and projects are captured by the AI Gateway in proxy and companion mode. For business users outside development, CompanyGPT is the right harness; our AI harness overview gives an overview.

From developer harness to enterprise harness

Coding harnesses are tools for developers. Business users, cost control and governance need two more layers.

CompanyGPT – the harness for business users

CompanyGPT gives any language model tools via MCP servers, integrations and n8n, company knowledge via companyRAG, roles from Entra ID or Keycloak, and audit logs – operated in your own cloud tenant (Azure, AWS or Google Cloud) or sovereignly on STACKIT.

Discover CompanyGPT
AI Gateway – cost and governance for every agent

The AI Gateway bundles the model access of your tools and coding agents behind an OpenAI-compatible API: sign-in via Entra ID, Keycloak or Cognito, budgets and hard limits per team and user, cost attribution per project and repository via the Companion, and logging of every request.

View the AI Gateway

Frequently Asked Questions

What is Paradev?

Paradev, called Superset until October 2026, is an open-source desktop app by Superset Inc. for running AI coding agents in parallel. Every task gets its own workspace with a Git worktree, branch and terminal. Paradev launches Claude Code, Codex, OpenCode, Gemini CLI, Cursor Agent and other CLIs and brings diff review, pull requests, an in-app browser, a CLI, an iPhone app and cloud workspaces.

Is Paradev open source and free?

Yes. Since 4 November 2026, Paradev has been MIT-licensed; before that, the Elastic License 2.0 applied. The Free plan costs nothing and covers one user, local workspaces, the desktop app and the CLI. Pro costs USD 20 per user and month, or USD 16 billed yearly, and adds remote access, automations and the mobile app. Team, Ultra and Enterprise include cloud workspaces.

Which coding agents does Paradev support?

According to the documentation, 22 agents are built in, including Claude Code, Codex, OpenCode, Gemini CLI, Cursor Agent, GitHub Copilot CLI, Amp, Pi, Oh My Pi, Grok, Devin, Muse Code, Droid, Kimi Code, Antigravity and Kiro. Claude Code and Codex are integrated most deeply, with status, resume, subagents and account switching. Any other CLI can be added as a custom agent.

Does Paradev use my Claude or ChatGPT subscription?

Yes. Paradev launches the installed CLIs with their existing login and, according to its docs, does not route model calls through its own servers. For Claude Code and Codex, several accounts can be stored and set as default. The usage page shows subscription quotas and estimates token costs at API list prices from the local session logs; for subscription usage, it shows your cost as USD 0.

Can Paradev be used in a GDPR-compliant way?

That depends on plan and configuration. Locally, code and sessions stay on the machine according to the maker, and prompts go to the agents' model providers. Telemetry to PostHog cannot be turned off according to the source code, and remote access, automations and cloud workspaces run through services based in the US. This belongs in your records of processing; the agents' model providers need to be assessed separately.

Does Paradev work with innFactory's AI Gateway?

Yes, through the agents it launches. Paradev itself makes no model calls for agent work. In proxy mode, you set `ANTHROPIC_BASE_URL` for Claude Code in the environment variables, while Codex uses its own configuration. The gateway's companion measures tokens and agent time from the local session logs, including subscription usage. For cloud workspaces, the local companion does not apply, because those sessions do not run on your machine.

Introduce coding agents safely in your company

We support selection, model access in the EU, permission design and cost control.