Traycer

Detailed report

Traycer keeps plans, agent work and reviews together across devices, with execution on computers you control. It supports headless hosts and phone access. Remote reliability and dependence on hosted coordination are the main issues to test.

Research checked November 13, 2026. Most documentation is undated. Dated releases, policies, and reports are identified below. This is document research, not a hands-on evaluation.

Product and workflow

The product is Traycer at traycer.ai, published by Traycer AI, Inc., with the official repository traycerai/traycer. The official site includes the Traycer mark alongside agent logos. The unrelated traycer.co describes an AI vector-graphics platform.

Traycer has also changed substantially: its IDE extension retired October 31, 2026. Older reviews of planning inside VS Code describe a predecessor, not the desktop product covered here. Official retirement notice

Traycer emphasizes continuity: keep the specification, tickets, conversations, agent work, and reviews together while changing models or devices. For an engineer already juggling Claude Code and Codex, that could remove substantial coordination work. The documented benefit is shared context and controls. This research does not measure coding quality. Product overview

Agents, models, and payment paths

The current compatibility table lists Claude Code, Codex, OpenCode, Cursor, Grok, Qwen Code, Kiro, Droid, Kimi, Copilot, Kilo Code, Amp, Devin, Pi, Hermes Agent, Oh My Pi, Reasonix, Antigravity, and Command Code, plus Traycer Inference, OpenRouter, and Hugging Face.

These integrations differ. Claude Code and Codex support account sign-in. Several others accept API keys. Cursor’s documented connection uses a Cursor API key. Models load dynamically from the selected integration and account, so there is no reliable universal model list. Terminal interfaces exist only for Claude Code, Codex, and OpenCode. The others use Traycer’s Chat interface. Permission support also varies: Cursor and Amp are listed as full-access-only. Agent compatibility and authentication

Your machines, headless servers, and reconnect

Desktop builds cover macOS Intel/Apple Silicon, Windows x64, and Linux. iOS and Android companions control agents running elsewhere. Phones do not host execution. Tablet support currently uses the portrait phone layout. Installers, mobile documentation

Crucially, your own headless server is a documented deployment. Install the CLI, run traycer login, approve its device-code login from another device, then run traycer host ensure. Linux requires systemd for the user service. The npm installation requires Node.js 22.13 or later. Homebrew is another documented route. Code and development tools must exist on the execution machine. Two-machine setup

SSH can provide the initial shell, but the documented ongoing connection uses Traycer hosts and its relay/account infrastructure. This is not merely an SSH terminal frontend. Port forwarding carries traffic over existing host connections, with loopback-only endpoints. Chat agents create the forwards. CLI server setup, port-forwarding design

Persistence has limits. Live agents, files, terminals, and Git state belong to their original host. Cloud sync carries Tasks and history between devices. If a host disappears, a Chat agent’s published copy can be read and cloned elsewhere, but files and uncommitted changes remain on the original machine. Terminal agents cannot move. Plain terminals close when their host goes away. This supports reconnect and continuation, not transparent process migration. Host ownership, offline behavior

Vendor cloud sandboxes are separately advertised as an extra-cost option. They should not be confused with registering your existing server. I did not establish their provisioning details or rates. Pricing

Isolation, orchestration, and shipping

Traycer supports Git worktrees, separate working directories and branches, for parallel changes. These prevent agents from editing the same checkout. They do not provide operating-system security isolation. Shared artifacts hold specifications, tickets, and reviews. Agents can create children, request reviews, and coordinate across your hosts. Overview

The communication matrix matters: all Chat agents can send and receive messages, but Terminal Codex/OpenCode agents cannot receive them. Agents cannot message another user’s agents, even in a shared Task. Cross-host acceptance means “queued,” not delivered. Current documentation specifies expiry after 24 hours. Agent-to-agent rules

Review includes per-turn changes, editable diffs, second-agent reviews, and rollback. GitHub PR panels use authenticated gh on the host to show checks and reviews and prepare feedback for an agent. Agents can run project tests, but this is not evidence of a guaranteed independent verification gate. Review workflow, PR integration

Price, privacy, and self-hosting

The published plans are:

• Free: individual use, unlimited agent sessions, cloud sync, device switching, remote hosts, mobile, and agent communication.
• Sync: $20/user/month monthly, or $16/user/month billed annually. Whole-team Task sharing requires Sync. Invitations to individual people work on Free.
• Inference: neither plan includes credits. One credit equals $1. Purchased credits do not expire. Traycer inference charges upstream API pricing plus 20%.
• External agent subscriptions/API charges remain separate, without Traycer’s inference markup. An important exception: Auto permission mode’s Automatic judge can consume Traycer credits even in external-agent conversations. Sandbox charges are additional. Detailed billing documentation

The repository is MIT licensed. Its contribution guide describes open-source clients, CLI, and protocol. I found no documented ready-to-use deployment for privately hosting the complete account, sync, and relay service. Hosting execution yourself is established. Fully self-hosting Traycer is not. Repository, contribution guide

The privacy policy, dated August 29, 2026, says Privacy Mode prevents logging/persistence of prompts, outputs, transcripts, and supplied code. It is default-on for teams and opt-in for individuals. Disabled mode permits prompt logging for 30 days. CLI-provider requests go directly to those providers, according to the README, which also discloses possible Sentry/PostHog telemetry. The policy’s no-persistence language needs clarification against documented cloud history sync. I would not infer that all synchronized content has the same protection as encrypted remote transport. Privacy policy, README data flow

Development and user reports

Development is active: GitHub lists Desktop/Host 1.5.3 on November 3, 2026, and 1.5.4 release candidates on November 12. That establishes maintenance, not operational stability. Releases

A concrete concern matches your use case: issue #2385, October 6, 2026, reported Mac-to-Windows agent messages remaining queued indefinitely despite working remote commands. A maintainer marked it fixed through an internal change on October 8. Treat it as a reported historical defect with a claimed fix, not a current confirmed failure. Issue and resolution

Independent praise is mostly older. An August 20, 2025 Reddit discussion valued verification against the original plan and responsive support. Critics argued the workflow could be recreated with existing agent tools. It concerns the extension and proves neither current desktop reliability nor consensus. Discussion

A LogRocket review dated October 23, 2026 reported useful security and intent findings but less review depth than Qodo. It mixes current desktop positioning with extension setup and obsolete paid-access claims, limiting its relevance. Vendor homepage testimonials are selected marketing evidence. Public independent desktop experience remains sparse. LogRocket evaluation

Decision and practical trial

Choose Traycer if you want several agents across your own always-on machines with desktop and phone control. Teams gain shared artifacts and visibility, but ownership restrictions limit agent collaboration between people. Avoid committing yet if you require air-gapped operation, proven unattended recovery, or fully private sync infrastructure.

The main concern is dependence on a young, cloud-mediated coordination layer for work that must survive disconnects.

Trial one real feature on a headless Linux server: use separate Claude/Codex worktrees, run existing tests, request cross-agent review, forward a development port, then switch laptops and steer from your phone. Disconnect the client, restart the host, and inspect history, delivery status, files, and billing. Success means reliable continuation with understandable failure states.