ComparisonFirst-party sources · reviewed 2026-09-18

T3 Code vs Skyflo

T3 Code is an open-source control plane for steering the coding agents you already use. Skyflo is the control plane for agentic engineering: one governed mission across those same agents, several repositories, independent review, and what the work teaches.

Direct answers on T3 Code plan mode, its orchestrator, and whether it is a harness, then the same eight dimensions every comparison uses, including where T3 Code is the better fit.

Every statement about T3 Code below is limited to behavior documented on the first-party pages listed at the foot of this page, on the date they were read. Where those pages do not cover a dimension, the table says so instead of guessing.

Evaluation brief08 dimensions

T3 Code · documented strength

T3 Code drives Codex, Claude Code, Cursor, Grok Build, OpenCode, and Antigravity from one MIT-licensed app on desktop, web, iOS, and Android, signed in with the subscriptions you already have. It adds durable threads, worktrees, checkpoints you can rewind, diffs, one-click pull requests, remote access, and usage limits across accounts.

Skyflo · where it sits

Skyflo makes the approved engineering objective the durable unit. One mission can span several registered repositories, gives bounded work to specialists, is checked by a reviewer that cannot edit, and leaves cited memory and Skills for later missions, with every change Skyflo makes to itself revertible.

Sources

27 first-party pages

Method

No winner score

Direct answers

T3 Code, from its own sources.

What is T3 Code?

T3 Code is a free, MIT-licensed app from T3 Tools for running the coding agents you already have. Its site calls it “the open-source control plane for coding agents”, and its README calls it an “agent harness control surface”.

It drives Codex, Claude Code, Cursor, Grok Build, OpenCode, and Google Antigravity through each tool's own sign-in, from a desktop app for macOS, Windows, and Linux, a web app, iOS and Android apps, and a CLI. It does not resell tokens. The latest stable release is v0.0.42, published on 16 September 2026, with nightly builds several times a day.

SourcesT3 CodeT3 Code READMET3 Code v0.0.42 releaseTerms of service

Does T3 Code have plan mode?

Yes, as a legacy setting that is off by default. On 7 August 2026 T3 Code removed the Build/Plan toggle from its composer, after Theo reported plan mode had fallen from 9% to 2.5% of sessions. Turning on Plan mode under Settings, General, Legacy features restores Build/Plan, the /plan and /default commands, and Shift+Tab. With it off, every thread runs in build mode.

T3 Code's plan mode is each provider's own planning behavior rather than a separate planner: Claude runs in its plan permission mode, Codex receives its plan collaboration mode, OpenCode uses its plan agent, and Cursor selects its plan mode. Antigravity is the exception: T3 Code's plan mode is unavailable there, and its docs point to Antigravity's own /plan command. The plan sidebar that used to open automatically became a folded summary inline in the conversation.

In Skyflo

The plan is a mission state rather than a mode. Discovery starts read-only, and implementation stays blocked until you approve the boundary the plan sets.

SourcesSettings contract (legacy plan mode)Plan toggle removed (#5551)Legacy features setting (#5664)Inline plan summary (#5558)Theo on plan mode usageTheo on legacy plan modeProvider adapters (plan mode mapping)Antigravity provider docs

How a Skyflo mission plans and asks for approval

What is the T3 Code orchestrator, and is orchestrator v2 released?

T3 Code's shipped orchestration is an event-sourced engine in its server. Commands are serialized, a decider emits events, and the event log is the source of truth for orchestration state. The stable release does not document an orchestrator that breaks an objective down and assigns it to several agents.

Orchestrator v2 is not released. It is pull request #2829, opened on 27 May 2026 and still unmerged on 18 September 2026. Its description covers a new runtime with checkpoints, rollback, and context handoff, new adapters for each provider, delegation tools, scheduled tasks, and subagent tracking. Builds that include it are labeled maintainer test builds and are not offered as updates. On 25 August Theo said mobile steering, queued sends in the background, and switching models in one conversation “kind of come with orchestrator v2”.

In Skyflo

Skyflo's orchestrator works on the mission rather than the thread. It gives bounded work to specialists, each coder in its own worktree, can pick the lead agent and model itself, and can change that routing policy only when a new one beats the current policy on past routes it was not built from.

SourcesOrchestration architectureOrchestrator v2 (#2829, open)Theo on orchestrator v2T3 Code v0.0.42 release

Is T3 Code an agent harness?

T3 Code presents itself as the surface over harnesses rather than as one. Its README calls it an “agent harness control surface”, and its site asks you to “plug in the harness you already pay for”: Claude Code, Codex, and the others are the harnesses. Each agent's loop, tools, and approvals stay inside that provider; T3 Code gives them one interface, durable threads, worktrees, checkpoints, and remote access.

In Skyflo

An engineering harness, as Skyflo uses the term, is the layer that holds one objective together across agent runs: the approved plan, delegation, execution surfaces, review, and memory between runs. Skyflo is built as that layer, and since Desktop 1.1.0 it is also a HyperAgent: it learns from the missions it completes, inside boundaries it cannot change.

SourcesT3 Code READMET3 Code

Which coding agents does T3 Code support?

Six: Codex, Claude Code, Cursor, Grok Build, OpenCode, and Google Antigravity. Each signs in with its own login, and Codex and Claude Code can run several accounts side by side. Antigravity became the sixth provider on 3 September 2026 and is off until you enable it. Claude Code runs through Anthropic's Agent SDK using your existing Claude Code sign-in.

In Skyflo

Skyflo runs Codex, Claude Code, Cursor Agent, Antigravity, Grok CLI, OpenCode, and Pi inside a Skyflo mission, alongside your own model keys held in the macOS Keychain.

SourcesT3 Code READMEProvider installationProvider architectureAntigravity provider (#9348)Claude Code accountsCodex accounts

T3 Code vs Skyflo: what is the difference?

They sit at different layers. T3 Code is where you steer individual agent threads, from any device. Skyflo is where one engineering objective lives: its approved plan, the specialists working across one or more repositories, an independent review, and what the work leaves behind.

What follows from that: a Skyflo mission can span an ordered set of repositories, its final reviewer cannot edit, completed missions leave cited memory, and a workflow proven in two repositories can become a Skill. Every change Skyflo makes to itself is listed in an Improvements ledger where one Revert undoes it. T3 Code is free and open source, runs on more platforms, and is the lighter choice when a single agent thread is the unit of work.

SourcesT3 CodeT3 Code READMEPermission modesRemote access

The fixed rubric

Skyflo vs T3 Code, dimension by dimension

The same 8 dimensions run on every comparison page, including the ones where T3 Code is stronger. A row reading “not addressed” means the reviewed documentation does not cover it, which is not the same as the product lacking it.

Dimension	Skyflo	T3 Code
01Primary unit of work	A persistent mission: one objective with its approved plan, work, review, and accepted memory.	A durable, event-sourced thread bound to one provider, with messages, activity, checkpoints, diffs, and session state.
02Execution environment	Skyflo Desktop runs against registered checkouts and credentials on a linked Mac.	The selected agent CLI runs on the machine hosting the T3 Code server; desktop, web, iOS, and Android clients reach it locally, through T3 Connect, by LAN pairing, or over desktop-managed SSH.
03Agent and model choice	An orchestrator assigns bounded work to specialists across supported signed-in harnesses and configured providers. Auto picks the lead agent and model per mission, and Skyflo can change that routing only when a new policy beats the current one on past routes.	Codex, Claude Code, Cursor, Grok Build, OpenCode, and Antigravity, each signed in with its own login or subscription. A thread stays bound to its provider, and several accounts per provider are supported.
04Cross-repository coordination	One mission can span an ordered set of registered Git checkouts under one approved plan.	Each thread runs in a fresh worktree or the current checkout of one project; the reviewed sources do not define one approved objective spanning several repositories.
05Approval and mutation boundaries	Discovery starts read-only, and plan mode blocks implementation until the user approves the boundary. External runtimes start at maximum-ask, and one Skyflo policy decides who answers each request and records the decision.	Supervised, Auto-accept edits, Auto, and Full access map onto each provider's own approvals, and Full access is the initial default. Plan mode is a legacy setting that is off by default.
06Independent review	A separate reviewer profile can read, search, and report but cannot edit or run mutating capability operations.	Diffs, review comments sent back to the agent, rewind with or without file changes, and pull-request workflows. Auto relies on the provider's automatic review; the reviewed sources do not define a separate reviewer without edit authority.
07Browser, terminal, and automation evidence	Code, browser, terminal, monitors, and automations stay attached to the same mission record.	Terminals, previews, diffs, source control across six Git hosts, usage limits, and remote clients are documented around the active environment and its threads. Scheduled tasks exist only in the unmerged orchestrator v2.
08Cross-mission memory, provenance, and user control	Source-linked memory Skyflo learns from completed missions under evidence rules. You can keep, correct, dismiss, or remove it, and revert Skills and routing changes from an Improvements ledger.	Threads, checkpoints, drafts, and provider session state persist, and provider skills can be invoked from the composer; the reviewed sources do not document memory learned in one thread and carried into another.

The decision

Choose for the shape of the work.

Choose T3 Code when

You want a free, MIT-licensed interface over the coding-agent CLIs you already pay for; T3 Code does not resell tokens.
Steering agents from desktop, web, iOS, Android, a paired device, or an SSH host is the experience you care about most.
One agent thread in one checkout is the unit of work, and each provider's own plan mode and subagents are enough orchestration.

Choose Skyflo when

One approved objective must coordinate changes across several registered repositories.
Plan approval and a final reviewer with no mutation authority must be product states, not provider settings.
What a finished mission proved should become cited memory, and a workflow proven in two repositories should become a reusable Skill.
You want Skyflo to pick the lead agent and model per mission, and to record every change it makes to itself where one Revert undoes it.

You might use both when

Use T3 Code to steer individual agent threads from any device, and Skyflo for the mission that coordinates repositories, approval, independent review, and what the work should teach the next mission.

Evidence ledger

First-party sources, dated.

Last reviewed 2026-09-18

Compare with other AI coding agents

Skyflo vs Claude Code Skyflo vs Cursor Skyflo vs OpenAI Codex Skyflo vs Devin Skyflo vs Factory Skyflo vs Conductor Skyflo vs Warp Skyflo vs Cline Skyflo vs Omnigent All comparisons

Evaluate with your own objective

Try both on the same objective.

Pick the piece of work you were going to evaluate on. We stay inside behavior available in the current Desktop build and label anything Preview or Planned as we go.

Download for Mac Book a current-build demo

Free. macOS 14+. Apple silicon and Intel.