Trebell. Download 2.1.6
New in 2.0: Muse Code and OpenCode 2

Trebell Code 2.0

A free desktop app for running AI coding agents on your own code. Use Trebell Native, our own agent, or Codex, Claude Code, OpenCode, Cursor, Grok Build, Antigravity and Muse Code. Each one runs as its real harness, with its own models, sign-in and approvals, next to your files, terminal, Git and pull requests, previews and a browser panel.

Version 2.1.6
Windows 10 and 11 (x64) · Release notes
Free
No cost to download or use
8 harnesses
Trebell Native plus seven coding agents
Trebell
Founder-led · founded 2026

The first major release since 1.6

2.0.0 is mostly about the coding agents you run inside Trebell. Muse Code is new, OpenCode 2 is supported, and every harness now brings its own models, approvals, sign-in and updates. We also redesigned the interface, made every turn show what it did, and made the app lighter on your machine. 2.0.1 then gives Trebell Native the same settings in the app as in our benchmarks. 2.1.0 adds Trebell Native on Meta's Model API and xAI, asks which model to use when a provider names no default, and lets each coding agent run on its own sign-in or its vendor's API key. 2.1.1 streams Trebell Native's requests to Meta's Model API, so a long answer is no longer cut off after five minutes. 2.1.2 keeps Trebell Native's turns on Meta's Model API going when a step needs one particular tool called. 2.1.3 takes longer messages and larger documents, shows each coding agent's own approval choices, and keeps a coding agent's quiet background work going for up to eight hours. 2.1.4 shows an estimated cost for Trebell Native's turns once you save your model's prices, counts cached tokens once in every estimate, and lets Trebell Native's checks of its own work wait for your approval. 2.1.5 lets Trebell Native read every document you attach at every access level and keep its conversations going after a restart, says why a declined step did not run, and keeps each conversation's own draft and files. 2.1.6 gives Trebell Native a Speed option on Anthropic as on OpenAI, with Anthropic's fast mode on Claude Opus 5.5, Opus 5 and Opus 4.8.

New harness

Muse Code, Meta's coding agent

Meta's coding agent now runs inside Trebell. Sign in with muse login, and its replies, reasoning, commands and file edits stream into the thread. Its approvals and questions come to Trebell's own card and dialog, and Usage shows its 5-hour and weekly windows.

Every harness

Each harness brings its own models

The model picker shows each harness's own models, options and default, and a reopened thread resumes on the model it used. Thinking shows as a collapsed Thinking row, OpenCode 2 runs beside OpenCode 1.x, and Settings offers an update when a harness has a newer release.

Design

A redesigned interface

Every page now shares the same top bar. Replies render as GitHub-style Markdown, with tables and code blocks you can copy.

Activity

Every turn shows what it did

Each turn keeps an activity card that says what every step did, such as a command's exit code and duration or an edit's diff. A turn that failed or was stopped before it replied says how and why.

Trebell Native

Native streams replies and waits out network drops

On OpenAI, Native's replies stream in as they are written, and code-change turns are checked before the answer. Native can look at images in your workspace, and it waits for the network instead of failing the turn.

Performance

Lighter on your machine

Trebell starts fewer background processes, uses less CPU and GPU while you type and replies stream, and opens large threads with much less disk and database work.

Version2.1.6
PlatformWindows 10 and 11 (x64)
PriceFree
UpdatesYou choose when to install

A look at Trebell Code 2.0

Projects and environments, files, a real terminal, Git and pull requests, worktrees, previews and a browser panel, around whichever coding agent you use.

Trebell Code 2.0 in dark mode: a Trebell Native thread in a demo project. The turn's activity card lists each step, the files it read and edited and a command it ran, with one edit's diff open, above the reply, and the Diff panel shows the three changed files. Trebell Code 2.0 in light mode: a Trebell Native thread in a demo project. The turn's activity card lists each step, the files it read and edited and a command it ran, with one edit's diff open, above the reply, and the Diff panel shows the three changed files.
Trebell Native: the turn's activity card lists every step, with an edit's diff open, above the reply; the Diff panel shows the changed files. Actual Trebell Code 2.0 interface, shown with demo data.

Eight harnesses to choose from

Trebell Native is our own agent harness. The other seven run as their real harnesses, each with its own models, sign-in, approvals and updates, so switching agents does not mean switching tools.

Trebell Native

Trebell's own agent harness: its own model and tool loop, calling model APIs directly, with OpenAI as the default provider. On OpenAI it streams replies as they are written and checks code-change turns before it answers.

own agent loopOpenAI defaultsees imageswaits out drops

Codex

Runs the Codex CLI you have installed, with its own sign-in and models. An approval card can allow commands or edits for the whole session instead of asking each time.

your Codex CLIsession approvalsgoals

Claude Code

Claude Code's own executable, with its own sign-in and models. Claude's to-do list shows as the thread's plan, and the context meter shows how full the context window is.

plan & to-doscontext meterBedrock · Vertex AI

OpenCode

OpenCode 1.x and, new in 2.0, OpenCode 2 through its own client and protocol. Its questions open in Trebell's dialog, and its turns survive a dropped event stream.

OpenCode 1.xOpenCode 2question dialog

Cursor

Cursor's agent, found at its official install location, with Cursor's own sign-in and its usage limits in Usage. It reads and searches files without asking; writes and commands follow your mode.

own sign-inusage limits

Grok Build

Grok Build signs in once with its own method and carries on, and Usage shows its meter. It reads and searches files without asking; writes and commands follow your mode.

own sign-inusage meter

Antigravity

Runs with a Gemini API key from Google AI Studio, kept in an owner-only file and never shown again. Attach images, audio and PDFs to a chat.

Gemini API keyimages · audio · PDF

Muse Code

Meta's coding agent, new in 2.0. Sign in with muse login in a Trebell terminal; its replies, reasoning, commands and file edits stream into the thread, and Usage shows its 5-hour and weekly windows.

new in 2.0muse login5-hour & weekly usage
Each agent keeps
its own modelsits own sign-inits own approvalsits own updates, from Settings
Settings, Agents & models in Trebell Code 2.0: the Agent harness card lists Trebell Native, built in and active, then Codex, Claude Code, OpenCode, Cursor, Grok Build, Antigravity and Muse Code, each with a Switch button.
Settings › Agents & models lists all eight harnesses: Trebell Native, built in, and seven more, each with a Switch button. Actual Trebell Code 2.0 interface, shown with demo data.

Why we built our own agent

The same model can do very different work depending on the harness around it: the tools it gets, the context it sees, how it recovers when something fails and how it checks its work. Trebell Native is the harness we build and tune ourselves, aimed at getting real repository work done reliably.

When the network drops

Long agent runs hit dropped connections and rate limits. If the harness gives up on the turn, the work done so far is lost.

  • New in 2.0: when a model request cannot connect, Native waits for the network and tries again every 60 seconds until it gets through or you stop the turn.
  • New in 2.0: on a rate limit, Native waits as long as the provider suggests, and with OpenAI reasoning models a dropped connection no longer loses the turn's reasoning.

Your real toolchain

An agent's commands run in your actual environment: your files, terminal, Git, worktrees, credentials and processes, with each provider's own quirks on top.

  • New in 2.0: commands Native runs keep your project's toolchain variables, so Ruby, Rust, Java, Go and Python builds find their tools, locally and on remote hosts.
  • New in 2.0: toolchain values that look like a password, secret, token or key are kept out of the commands Native runs.

Checking the work

An agent saying it is done does not mean the code works. Native checks before it answers.

  • New in 2.0: Native recognizes more ways of asking for a code change and checks its work on those tasks before it answers.
  • New in 2.0: when its check cannot confirm the work, Native says why, and keeps or restores the version of your files the evidence best supports.

How Native finds the right code

More context is not always better. Trebell Native reads the files that define a name first, searches beyond source code, and keeps long output out of the model's way without throwing it away.

  • Definitions firstThe files that define a name are read before the files that only use it, and results come back grouped by file.
  • Beyond source filesCode search also finds matches in docs, templates, configuration and changelogs, and a search can be limited to a folder or a glob.
  • Long output, keptOn long tasks, runs of identical output lines are collapsed for the model, with the full output still retrievable.

What a finished task costs

We compare cost on finished work, task by task. On fresh tasks where Trebell Native and Codex using the OpenAI API got the same verifier result, Native's API-equivalent cost was 18% lower on SWE-bench Multilingual and 28% lower on Terminal-Bench 2.0. On Terminal-Bench Pro 1.0 the cost was statistically level.

  • Same model & settingsgpt-6-luna with max reasoning on the fast service tier for both harnesses, with hosted web search off.
  • Same result firstCost is compared only on the tasks where both harnesses got the same verifier result.
  • 95% intervals0.75 to 0.90 on SWE-bench Multilingual and 0.58 to 0.87 on Terminal-Bench 2.0, from bootstrap resampling.

How Native compares with Codex

Trebell Native and Codex using the OpenAI API ran the same model and settings, gpt-6-luna with max reasoning on the fast service tier and hosted web search off, on fresh tasks that had never been used before. An independent verifier graded every run. Cost is API-equivalent, compared on the tasks where both got the same verifier result.

SWE-bench Multilingual · 156 fresh tasks 18% lower cost
125 vs 124

One more task resolved, for less

Trebell Native resolved 125 of 156 tasks and Codex using the OpenAI API resolved 124. On the tasks where both got the same result, Native cost 0.819x as much, with a 95% interval of 0.75 to 0.90.

Whole fresh pool · tasks resolved
Terminal-Bench 2.0 · 78 fresh tasks 28% lower cost
62 = 62

Same result, 28% less spend

Trebell Native and Codex using the OpenAI API each resolved 62 of 78 tasks. On the tasks with the same result, Native cost 0.721x as much, with a 95% interval of 0.58 to 0.87.

Whole fresh pool · tasks resolved
Terminal-Bench Pro 1.0 · 40 fresh tasks 1 more resolved
22 vs 21

One more task resolved

Trebell Native resolved 22 of 40 fresh tasks and Codex using the OpenAI API resolved 21. Cost on this pool was statistically level: 0.929x, with a 95% interval of 0.80 to 1.08.

40 fresh tasks · tasks resolved
BenchmarksSWE-bench Multilingual · Terminal-Bench 2.0 · Terminal-Bench Pro 1.0
Model and effortgpt-6-luna · max reasoning · fast tier
ComparedTrebell Native · Codex (OpenAI API)
GradingIndependent verifier · fresh tasks

About Trebell

Trebell is a developer-tools startup founded in 2026. Its main product is Trebell Code, a free desktop workspace for AI coding agents, now at version 2.0. The company is founder-led, with a five-person core team.

Founder and team

Tanishq Baweja, Founder & CEO

Tanishq founded Trebell and currently leads product, engineering and company strategy. The core team also includes Raghav Dabra (Engineering), Himnish Ishaan (Product & Research), Shashwat Chaudhary (Operations & Growth) and Samriddhi Choursarisa (Engineering). Full company & team profile →

Our agent

Trebell Native is our own agent runtime and harness

It runs its own model and tool loop and calls model APIs directly, with its own code search, permission modes, recovery and checks, instead of passing everything through another coding harness.

Cost

Lower cost for the same work

Good repository search, long output kept out of the prompt, context that stays stable for the cache, and sensible stopping and retries all cut cost without needing a smarter model.

Today

Trebell Code 2.0 is out now

Trebell Code 2.0 ships Trebell Native and runs seven more coding agents in one workspace with files, a terminal, Git and pull requests, worktrees, previews, a browser panel, projects and environments. Trebell Native's benchmark results are graded by an independent verifier.

Next

More fresh benchmarks

Today's results come from 274 fresh tasks across SWE-bench Multilingual, Terminal-Bench 2.0 and Terminal-Bench Pro 1.0. We want gains that hold up on new tasks, not one good benchmark run.

Providers

Not tied to one model vendor

Trebell Native calls model providers directly, with OpenAI as the default, and Trebell Code also runs seven other coding agents as their real harnesses. Which model you use stays your choice.

Teams

A paid plan for teams, later

The planned commercial layer adds organization policy, shared usage and cost controls, managed execution environments, auditing and administration, without changing how developers work.

Business model

Free for individual developers

Trebell Code is free, and you can start using it today. We plan to charge engineering teams for the shared layer: organization policy, shared usage and cost controls, managed execution environments, auditing and administration.

Get Trebell Code 2.0

Trebell Code is free and runs on Windows 10 and 11 (x64). Use it with Trebell Native or with the coding agents you already have.

Download for Windows (x64)

Trebell Code 2.1.6 for Windows 10 and 11 (x64), free to download and use. Release notes. The installer is not code-signed yet, so Windows SmartScreen may ask you to confirm: choose More info, then Run anyway.

About Trebell & team →