T3 Code vs Caprock: two open-source apps for running coding agents, compared

T3 Code is a cockpit: a free, MIT-licensed app for macOS, Windows and Linux that starts Claude Code, Codex and six other agents in its own chat view, over a server you can host on another machine, with native phone apps. Its Usage screen reads every provider's history from disk. Caprock is the ledger: every agent session on the machine, live, whether or not it started it, priced at list price and joined to the pull request, with each agent in a real terminal tab and no telemetry. T3 Code's cells were read 11 October 2026; its V2, phone app and telemetry re-read 24 October, its FAQ, Usage docs, stars and release v0.1.3 on 27 October 2026.

What T3 Code is

T3 Code calls itself an "agent harness control surface". A server runs on the machine where the agents work, and you drive them from its Electron desktop app, a web app or its iOS and Android apps. It works with your own logins to Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity and Pi, Muse Code in beta, and agents from the ACP Registry. It is MIT-licensed, from T3 Tools Inc., and its README is frank about the stage: "We are very very early in this project. Expect bugs." Its FAQ of 26 October 2026 is as plain about money: "There's no way to pay us to use T3 Code", and T3 Connect, its relay, is "free too, for now … if the bill gets unacceptable we may have to change course". It had 30,749 GitHub stars on 27 October 2026.

Its newest stable release on 11 October 2026 was v0.0.45. Its Orchestrator V2 shipped as v0.1.0 on 15 October 2026, and v0.1.3 on 27 October added guest sharing of a thread, pull request management from the phone and a link to "pull requests agents create from the shell". The move to V2 copies the V1 database once; titles, messages, worktrees and linked pull requests come across, old checkpoints and diffs, tool activity, approval history and proposed plans do not, and new conversations stay in the version that made them (its issue #14871).

T3 Code vs Caprock, side by side

A server you host, or your other computers

T3 Code's server can run on a remote machine. You open your instance in a browser or connect the desktop app, approve with a short code, and every agent runs on that server; the phone and the desktop are windows onto it (its remote-access docs, read 24 October 2026). A user who self-hosts it put it this way on 22 October 2026 (translated): "open my instance, enter my code, use it like a desktop app, all models on the server."

Caprock runs on each computer. Your other computers that run Caprock link to this one through its encrypted relay, and the sidebar lists the sessions running on each, one computer at a time: open one, read its conversation, type into its terminal. Starting a session on another computer from the app comes next; one view that adds the computers up is not there yet. Nothing in it runs on a server of ours.

A chat view or the agent's own terminal

This is the real difference. In T3 Code you write in its composer and read the reply in a thread; approvals, diffs and file restores are T3 Code's own screens, and a terminal panel sits beside the thread. A thread can switch provider and carry its context with it.

In Caprock each agent runs its own interface in a real terminal tab, the same screen you get in your shell: its slash commands, its prompts, its keys. Caprock adds around it the tabs per project, the sessions waiting on a permission prompt first, and the cost; each agent's tab also switches to a chat view of the same session. The Claude Code you start in your own terminal shows up beside them, live, with what it cost.

What each one sees: threads it started, or every session

T3 Code is the cockpit closest to a ledger, because its Usage screen reads from disk: it "combines Codex, Claude Code, Grok Build, OpenCode, Antigravity, Pi, Muse Code, and Cursor history from your connected environments", so a session you started in your own terminal is counted in its totals (its usage docs, read 27 October 2026). What it does not do is watch one live, or join it to a pull request: a thread is live only if T3 Code started it, and importing an outside session as a thread exists but its open issue #10961 says "externally started sessions can never be imported" after onboarding. Its V2 shows "no token usage for subagents" (its issue #15429, open).

Caprock records the machine, not its own tabs:

• Sessions it did not start, live. A Claude Code session running in your own terminal shows up live, with its cost; Codex, OpenCode, DeepSeek Harness and Hermes Agent come in from the history they write.
• Each pull request, with what it cost. The Pipeline shows each pull request with its checks per job and OS and what the agents spent on it; T3 Code links a thread to its pull request and checks and shows cost per model and per account.
• Bedrock and Vertex. Caprock prices a session by its model id, and strips the Bedrock (us.anthropic.…) and Vertex (…@date) forms to the same name before it looks up the price (internal/cost/cost.go), so a Bedrock session costs what the same turns would cost on the API.
• Not the cloud, not CI. Sessions that run in a vendor's cloud or inside CI are not observed; Caprock sees the machines it is installed on.

The phone: T3 Connect, direct pairing or an encrypted relay

• T3 Code has native apps on the App Store (version 2.1.0 on 24 October 2026) and Google Play. A phone reaches your machine through T3 Connect, which needs a sign-in, or by a pairing link over your LAN or Tailscale. Push alerts when an agent finishes or needs approval go through T3 Connect; a direct or Tailscale connection alone does not deliver them.
• Caprock has no phone app. The phone opens Caprock in its browser, paired by a QR code, with no account, from anywhere through a relay that carries only end-to-end ciphertext. It can control sessions unless you untick that when you connect it. Alerts can also reach Telegram through a bot you create.

Cost, plan limits and telemetry

• T3 Code's Usage screen estimates API-equivalent cost, tokens and cache savings per model for eight providers, across every machine you connect, labelled not your subscription bill, and pools the plan windows of all your Claude or Codex accounts into one bar per window. On the limits themselves, one issue read "Usage is ~3x-5x higher than the native Claude app" (#7338, 71 comments, closed), and its FAQ now says that T3 Code drives Claude through Anthropic's Agent SDK and "all Agent SDK usage will drain your 5h limits faster than using Claude Code directly" (read 27 October 2026).
• Caprock prices every turn of Claude Code, Codex, OpenCode, Gemini CLI and DeepSeek Harness per repository and per pull request, and shows the Claude and Codex plan windows, one per sign-in, with a warning before a limit. Every dollar Caprock shows is an estimate at API list price, as Claude Code's own /usage is: not a bill, and on a flat plan not money saved.
• Telemetry: T3 Code's server sends product usage events (provider, model, permission mode, turn result, duration, token totals; never prompts or code) to ingest.t3.codes, which forwards them to PostHog, with hashed ids for your T3 Connect, OpenAI and Claude accounts and one per device; T3 Connect's connection traces go to Axiom. Its docs give one switch, T3CODE_TELEMETRY_ENABLED=false in the server's environment; a setting in the app is not in them. Its issue #4123, on telemetry being on by default, was reopened by the maintainer on 22 October 2026 (read 24 October 2026). Caprock has none.

Permission prompts by default

T3 Code's initial default for new threads is Full access: commands and edits run without approval prompts until you pick Supervised, Auto-accept edits or Auto in Settings. Caprock starts new sessions accepting edits and asking before a command; bypass is one pick away in the New agent sheet.

T3 Code on Windows

T3 Code's desktop app installs with winget install T3Tools.T3Code or Scoop, and the t3 command with a PowerShell script. It can run the agents inside a WSL distro. Its background service is documented for macOS and Linux only. Caprock ships a Windows installer; it is not code-signed yet, so SmartScreen asks for Run anyway once.

Download size

Measured on 11 October 2026 from the files each project publishes on GitHub: Caprock v0.100.0 and T3 Code v0.0.45. T3 Code ships Electron; Caprock uses the web view the system already has, except in the Linux AppImage, which carries its own WebKit.

T3 Code's macOS row is its Apple-chip .dmg and its Windows row the x64 .exe. These are sizes on disk; neither app's memory use or speed has been measured against the other.

T3 Code vs Orca, Conductor and Claude Code

• Orca is an IDE around the agents: a code editor, a Chromium browser, SSH and phone apps; see Orca vs Caprock.
• Conductor is a Mac cockpit for the same job: free for local workspaces; Pro, $50 a person a month, adds cloud workspaces and the iOS app; Teams, $60 a user a month, an admin portal and central billing; cost per person or per repository is not in its docs (its pricing, read 27 October 2026). Compared with the others in desktop apps for coding agents.
• Claude Code is the agent itself. T3 Code and Caprock both run the claude you have installed and signed in to; neither replaces it.

What Caprock does not do

• Not the cloud, not CI. Sessions that run in a vendor's cloud or inside CI are not observed; Caprock sees the machines it is installed on.
• Not a bill. Every dollar Caprock shows is an estimate at API list price, as Claude Code's own /usage is: not a bill, and on a flat plan not money saved.
• No code editor. Files open read-only in a tab; you edit in your own editor.
• No SSH remotes, no cloud machines. It reaches only your other computers that run Caprock, through its relay.
• No native phone app. The phone uses its browser, which can be added to the home screen.
• No code-signed Windows installer yet. First launch on Windows needs SmartScreen's Run anyway; the Mac app is notarized.
• No cost for Grok or custom agents. Caprock starts them but does not read their turns.

Which to pick

• T3 Code — you want a chat-style cockpit with native phone apps, a server you host and open from anywhere, several machines in one list, or Cursor and Antigravity beside Claude Code and Codex.
• Caprock — you want the ledger: every session on the machine, live, the ones from your own shell included; who is waiting on you; each limit; and what each pull request cost, on the API, Bedrock or Vertex alike, with each agent in its own terminal and no telemetry. Install it.