TermLoop

Native agent operations

Keep every coding agent in the loop.

TermLoop puts tasks, worktrees, real terminals, diffs, and agent coordination in one native workspace—without replacing Claude Code or Codex.

Available nowOpen sourceYour subscriptionsNo telemetry
launchpad · live workspace
Real app. Real terminals. Real state.17 seconds · no concept footage
Bring your agentsClaude Code · Codex
Keep real GitBranches · worktrees · diffs
See every terminalGhostty-powered PTYs
Work from anywhereDesktop · mobile companion

The TermLoop model

Simple objects. Serious workflows.

Projects hold the durable scope, Tasks describe work, and Sessions are the live agents and terminals operating inside that Project.

01 / PROJECT

One durable scope

A repository or folder, its connections, tasks, sessions, and launch configuration—kept together without hiding the filesystem.

project → tasks + sessions
02 / TASK

Work with optional isolation

A Task can stay lightweight or get one managed Git worktree when the work needs its own branch and checkout.

task → 0..1 worktree
03 / SESSION

Every process stays visible

Claude, Codex, terminals, and dev runs stay Project-scoped. Select them, refresh them, fork context, or coordinate them directly.

session → agent | terminal | run

Product tour

From prompt to reviewed change.

Every demo below was recorded in the real desktop app with real TermLoop Sessions. Use the index like documentation, or scroll through the workflow end to end.

Run

01 — launch and isolate work
01 · Quick Actions

Start the right agent with the right context.

Open the command palette, compose a free prompt, choose the agent, model, permissions, and reasoning, then launch an inspected Session in seconds.

Shift ShiftVersioned promptClaude / Codex
quick action · inspected launch
02 · Task worktrees

Isolation when the task needs it.

Create a Task first, then provision one managed worktree when implementation should happen on its own branch. The Task stays useful even without a checkout.

Real GitOptional worktreeSafe cleanup
task · managed worktree
03 · Dev server

Define it once. Start it where you work.

Save a Project dev-server command and fallback URL, then launch the real process from the Project or its Task worktrees. Its terminal and discovered URLs stay visible.

Live processProject configPreview URLs
launchpad preview · :4173

Coordinate

02 — agents working together
04 · Multi-agent

One workspace. Several live agents.

Run multiple Claude and Codex Sessions across tasks and worktrees. See who is active, which checkout changed, and which terminal needs attention.

Parallel SessionsLive statusShared Project
active agents · launchpad
05 · Ask To

Get a second opinion without copy-paste.

Ask one agent to consult another provider. TermLoop creates a visible helper Session, preserves the relationship, and returns the answer to the requester.

MCPClaude ↔ CodexFollow-up ready
ask_to · release review
06 · Handoff

Move context to the agent already in the work.

Send a visible, composed handoff from one running Project Session to another. The target receives the brief in its own real terminal and continues from there.

Visible messageRunning SessionNo clipboard
handoff · activity feed
07 · Session fork

Explore another direction in a real terminal.

Fork a Session from the sidebar, see the parent-child relationship, and continue in a new Ghostty terminal. The branch in thinking is explicit instead of becoming another anonymous tab.

Real terminalVisible lineageIndependent context
session fork · onboarding

Review

03 — understand and control state
08 · Changes

Review the diff where the work happened.

Changed-file counts appear on Tasks and active Agents. Open the Git changes view, inspect the real diff, and leave line-level notes for the agent to address.

Git diffTask countsAgent notes
changes · onboarding checklist
09 · Task lifecycle

Close the work without erasing it.

Tasks have a deliberately small lifecycle: open or closed. Close finished work, keep its current state visible, and reopen it when the next change arrives.

Open / closedReopenState preserved
task · close and retain
10 · Session refresh

Refresh the display, not the identity.

Recover a stale agent display from the Session menu while keeping the same durable Session and process model. Lifecycle actions remain explicit.

Stable SessionDisplay recoveryExplicit lifecycle
session · refresh display

Operate

04 — supervise beyond one terminal
11 · Steward

A Project-scoped operator that stays present.

Ask the persistent Steward what should ship next, let it coordinate Tasks and Sessions through its constrained tools, and keep the conversation attached to the Project.

PersistentProject-scopedCapability-limited
steward · launchpad
12 · Mobile companion

Check the loop when you leave the desk.

Pair your phone, keep Project visibility and terminal input scoped to your computer, and follow long-running agents over your trusted network.

QR pairingTailscale readyLive terminal
mobile · secure pairing
13 · MCP & prompts

Inspect the tools and instructions agents receive.

Browse role-scoped MCP tools and the built-in prompt catalog. TermLoop-generated instructions remain visible and editable; provider-managed opaque layers are labeled honestly.

Role profilesPrompt catalogInspectable
application settings · MCP

Product principles

The terminal stays real. The workflow gets better.

01

No replacement agent

TermLoop launches the Claude and Codex CLIs you already trust and pay for.

02

No hidden filesystem

Branches, worktrees, processes, and files remain ordinary local resources you can inspect outside the app.

03

No mystery prompts

TermLoop-authored prompts are visible assets. Secret values stay out of durable state and public evidence.

04

No fake coordination

Agent-to-agent requests, handoffs, terminals, diffs, and lifecycle changes are real product actions.

Straight answers

Frequently asked.

Which agents does TermLoop support?

The first-class agent launch paths are Claude Code and OpenAI Codex, alongside ordinary terminal and dev-run Sessions. TermLoop uses your existing provider sign-ins rather than proxying model traffic.

Does every Task create a worktree?

No. A Task can have zero or one worktree. Create the Task to track work; provision a managed checkout only when branch isolation is useful.

Can multiple agents write in the same worktree?

Yes. TermLoop does not impose a product-level write lease or artificial launch gate. You choose whether agents share a checkout or work in isolated Task worktrees.

Where do my code and credentials go?

Your code remains on your filesystem. Agents connect directly to their providers. TermLoop keeps secrets out of durable product state, logs, URLs, and generated evidence.

Which desktop platforms are available?

The current public desktop download is for macOS. TermLoop Next treats macOS, Linux, and Windows as release gates; broader availability follows verified native builds rather than an untested compatibility claim.

Is the mobile connection real?

Yes. The companion pairs with a running desktop instance for Project visibility and terminal input. The public demo intentionally masks the one-time QR payload.

Is TermLoop open source?

Yes. The public source and releases are available on GitHub, including the terminal and agent-workflow layers.

More agents. Less terminal babysitting.

Download TermLoop and keep the whole engineering loop visible.