Tracer is a free open-source app that orchestrates Claude Code, Codex, Cursor and other coding agents in one workspace, on desktop and mobile.
What is Tracer? #
Tracer is a free, open-source desktop application for Mac, Linux, and Windows that gives coding agents like Claude Code, Codex, OpenCode, and Cursor a shared workspace. Instead of running each tool in its own window and copying context between them, Tracer lets the agents work side by side on the same task, talk to each other, and hand off work, while you direct everything from one interface. It uses your existing subscriptions rather than requiring a new paid plan to access the agents.
TL;DR #
- Bring-your-own-subscription model means Tracer doesn’t charge extra to run Claude Code, Codex, OpenCode, Cursor, or other agents since you connect your existing accounts.
- Tasks act as persistent containers holding chats, files, artifacts, git diffs, pull requests, terminals, and browser sessions, so switching models mid-conversation doesn’t lose context.
- Agents can collaborate directly , asking each other questions, requesting reviews, or handing off work without you manually relaying messages between chats.
- A remote host setup lets you run agents, terminals, and browser sessions on one machine (laptop, desktop, dev box, or VM) and control them from another device, keeping code and credentials on the host.
- Mobile apps for iOS and Android let you start tasks, answer agent questions, approve permissions, and use a full terminal (with an on-screen key bar for escape, tab, and arrow keys) from your phone.
- An in-app browser panel lets both you and the agent navigate, inspect, click, screenshot, and verify work, including a visual annotation tool for flagging UI issues.
- Model routing automatically falls back to another account or provider when one hits a rate limit, keeping the conversation alive without losing chat history.
How does Tracer organize multi-agent work? #
The core unit in Tracer is the “task.” A task bundles everything related to one piece of work: the chat history, generated files, artifacts, git diffs, pull requests, open terminals, and browser sessions. Because all of this lives in one place, you can switch the underlying model partway through a conversation (say, from Claude to GPT-based Codex) without losing the thread. That matters in practice because any single model can get stuck or loop on a problem, and being able to pull in a second model for a fresh take, without re-explaining the whole context, saves real time.
Panels for chat, files, diffs, terminals, and the browser can be arranged side by side, so you can watch an agent’s output and its effect on the running app at the same time rather than tabbing between windows.
Tracer also includes built-in “skills,” which are reusable workflows for common tasks like planning, debugging, code review, and documentation. Teams can use multiplayer workspaces to open shared tasks and comment on artifacts together, which is useful when more than one person needs visibility into what an agent is doing.
Why let agents talk to each other? #
A common friction point in using multiple AI coding tools is the manual copy-paste loop: you ask one model a question, take its answer, paste it into another model’s chat, and relay the response back. Tracer removes that step. Agents inside the same task can ask each other questions directly, request a review of a change, or hand off a piece of work to another agent without the human acting as a message courier.
This is particularly relevant for people who already pay for multiple coding assistants and want to use each one for what it’s best at, for example, one model for fast scaffolding and another for careful review, without the overhead of managing separate conversations.
What does the remote host setup actually do? #
Tracer’s remote host feature turns any machine, laptop, desktop, dev box, or VM, into the place where the real work happens: the repository, running agents, terminal sessions, browser sessions, and credentials all live there. You then connect from a second device (another computer or a phone) and that device becomes a window into the host rather than a place where code or secrets are copied.
Setup involves registering a host, then connecting from another device to pick up the same tasks, chats, artifacts, terminals, and browsers exactly where they were left. Access is tied to your Tracer account, and you can view and remove registered hosts and devices at any time, which matters if a phone or laptop is lost.
How does the mobile app change the workflow? #
Tracer ships native apps for iOS and Android that connect to a running remote host. This isn’t a passive dashboard for checking status. You can start a new task from your phone, including attaching an image for context, and the agent begins working on the host machine.
Enterprises have Remy.
When an agent needs input mid-task, such as answering a clarifying question or approving a permission, you can respond directly from the phone. Push notifications alert you when an agent needs attention or finishes a task, and tapping the notification opens that specific task. Panels (files, artifacts, diffs, pull requests, terminal, browser) open full screen on mobile and the browser panel streams from the host to the phone.
The terminal includes an on-screen key bar with escape, tab, control, and arrow keys, since phone keyboards lack them natively. That makes it genuinely workable to run terminal commands from a phone rather than a desktop simulation.
There are limits worth knowing: the phone requires an active remote host (agents don’t run locally on the device), the phone can’t act as a host itself or restart one, and voice dictation isn’t yet supported in the mobile app.
What can the in-app browser do? #
Tracer includes a browser panel inside each task, positioned alongside chat, terminal, and file panels, and both the user and the agent can control it. The browser’s start page automatically surfaces detected local dev servers, so if your app is running, you click through to it directly.
For visual bugs, instead of describing an issue in text, you can select and annotate the specific component on the page and send that crop as chat context to the agent, cutting down on back-and-forth clarification. Agents can also use the browser autonomously: navigating pages, inspecting elements, clicking through flows, and taking screenshots to verify that a change actually works. If the agent hits a login screen, CAPTCHA, or multi-factor prompt, it hands control back to the user to complete that step before resuming. The browser can also pop out into a floating picture-in-picture window so the running app stays visible while you work in other panels. One caveat: when the browser streams from a remote host to a separate desktop, you get basic navigation and the address bar, but local-only features like dev tools, viewport presets, and annotation aren’t available in that mode.
How does model routing handle rate limits? #
Anyone running agents against commercial model APIs eventually hits a rate limit or an outage mid-task. Tracer’s model routing feature is built to handle that without killing the conversation. In settings, you configure a fallback plan: try a different account on the same provider, switch to a comparable model from another provider, or wait for the original provider’s usage limit to reset. If none of those options work, Tracer notifies you.
You control the order of fallback steps, define which models are acceptable substitutes for each other, and can cancel a pending switch if you don’t want it to happen. You also choose whether the session automatically switches back to the original provider once it’s available again. By default, automatic routing is turned off and must be enabled manually. It applies only to chat agents on the selected host, not terminal agents, and when a switch happens, Tracer starts a new agent session that inherits the existing chat history so context isn’t lost. The wait-for-reset option only works if the provider actually reports a reset time, and only if that time falls within a maximum wait window you configure.
For anyone juggling multiple model subscriptions, this removes one of the more annoying interruptions in agent-based coding work.
What is Tracer building next? #
› Build a safety-inspection app for 60 plants.
already included:✓ SOC 2✓ role-based access✓ audit log✓ self-hosted
According to the creators, the next phase of development is long-running automations, agents that handle ongoing engineering work in the background and only surface to the user when a decision or approval is genuinely needed. The stated vision includes direct integration with observability and monitoring tools such as Sentry, PostHog, and Datadog, so automations can watch production systems without a human checking dashboards constantly. The roadmap also describes self-improving workflows that periodically review their own performance, cloud-based execution so automations aren’t tied to a laptop staying open, and proactive suggestions where the system looks at your codebase and existing workflows to recommend new things to automate.
This functionality is not yet released, it’s described as a roadmap item with a waitlist, and it remains provider-agnostic and bring-your-own-subscription, consistent with how the current desktop and mobile apps work.
Frequently Asked Questions #
Is Tracer free to use?
Yes. Tracer is described as a free and open-source application. It does not require a separate paid plan to use agents like Claude Code, Codex, or Cursor; instead, you connect your own existing subscriptions to those services.
Which coding agents work with Tracer?
The transcript names Claude Code, Codex, OpenCode, and Cursor as agents that can run side by side inside Tracer, with support for “other coding agents” as well.
Does Tracer store my code or credentials on my phone?
No. In the remote host setup, the host machine holds the repository, running agents, terminals, browser sessions, and credentials. A phone or second device only displays and interacts with that host; it does not copy code or credentials locally.
Can I use Tracer without a remote host?
The mobile app specifically requires an active remote host, since agents don’t run directly on the phone and the phone cannot act as a host or restart one. The desktop app can run locally without a separate host machine.
What happens if a model provider hits a rate limit mid-task?
If model routing is enabled, Tracer can automatically switch the chat to a fallback account, an equivalent model on another provider, or wait for the usage limit to reset, depending on the fallback order you configure. It will notify you if none of the configured fallbacks succeed.