Use open modelsin Claude Code

Use Claude Code exactly as you do today: same interface, same workflow. TogetherLink routes each task to the right model, so easy work stops burning expensive tokens and your monthly AI bill drops by about 80%.

Read the docs

macOS and Linux.

2
agents live, one coming
4
curated models
1
install command
MIT
open source

Type one command. That's the whole setup.

Real sessions, seconds each: the short command launches your own agent on an open model, and prints the cost when you leave.

Your own Claude Code on Auto. The Together lineup sits in /model.

Config untouched

~/.claude is never edited. The desktop profile is backed up first and comes off with one command.

Logins survive

Claude stays signed in, in the terminal and the app. Subscriptions stay yours.

Cost on exit

Proxied sessions print their token and dollar totals.

Trying it costs one paste.

Set the model to Auto.

The router weighs every task on its way through. Quick work rides GLM 5.3 Flash for pennies; the hard problems get Opus 5, and the receipt shows both.

Claude CodeModel · Auto
Fix the failing lint rule
glm-5.3-flash · $0.003
Routerbalanced
EasyHard
Weight 0.14Flash
Opus 5
The hard calls
3 routed$1.84
GLM 5.3 Flash
The easy calls
11 routed$0.04
Send one through:

Two agents live. One launcher.

Claude Code and Claude Desktop are wired today; ChatGPT Desktop is next in line. Each is one short command away, and nothing else on your machine changes.

Claude Code

beta

A local proxy translates Claude Code's traffic to Together. You stay signed in, and the session prints its cost on exit.

Claude Code
togetherlink
Together AI
What stays yours
  • Your Claude Code login and subscription
  • ~/.claude settings, memory, and history
  • Web search, via an optional Exa key

Claude Desktop

beta

One command installs a reversible profile and reopens the app with the Together lineup in its model menu, for Chat, Cowork, and Code alike. Backup first; one command puts it back.

Claude
togetherlink
Together AI
What stays yours
  • Your Claude login, projects, and chats
  • Every setting in the desktop app
  • The native models, in the same menu

ChatGPT Desktop

coming soon

Next on the board. One command will patch the desktop app's model picker, with a backup taken first and one command to restore it exactly.

ChatGPT
togetherlink
Together AI
What stays yours
  • A backup under ~/.togetherlink/backup/
  • Native GPT models, still in the picker
  • One command restores everything

Your Claude is on the board. One paste wires it up.

Four models, priced in the open.

Auto Router is the default; pick a model and an agent to compose the exact run.

Compose a runmodel x agent
$togetherlink --model moonshotai/Kimi-K3 claude

Launches Claude Code on Kimi K3.

A month of this, on the meter.

Drag your hours; the receipt reprints for Kimi K3.

2.0h
monthly receiptKimi K3
Open models, this month$34.24
Closed frontier API, same month$171.21
You keep$136.97

From install to routed in ninety seconds.

  1. Run the installer

    One paste. Everything lands in ~/.togetherlink and keeps itself current.

  2. Launch an agent

    tclaude for the terminal, tclaude-desktop for the app. First launch asks once for your Together API key.

  3. Work as usual

    Same tool, open model. The receipt prints what the session cost.

TogetherLink versus doing it by hand.

You could export the keys and edit every agent's config yourself. The difference shows up the day you want your old setup back.

TogetherLink
Your configUntouched
Applies forOne run
RoutingAutomatic, per launch
AnalyticsReceipt + usage report
By hand
Your configYou edit it
Applies forForever
RoutingNo routing
AnalyticsNone built in

Common questions, short answers.

Longer ones live in the official docs and the LLM docs, which your agent can read on its own.

Does it touch my agent config?

Claude Code runs on settings that exist only for the session; ~/.claude is never edited. Claude Desktop gets a reversible profile with a hash-checked backup: togetherlink claude-desktop off restores only the keys TogetherLink wrote. ChatGPT Desktop, when it lands, follows the same backup-first rule.

Do I need Claude installed first?

Yes. tclaude needs the Claude Code CLI; tclaude-desktop needs the Claude Desktop app. If either is missing, togetherlink prints the official install command and exits.

What does it cost?

TogetherLink is free and MIT licensed. You pay Together per token, at the prices above; proxied runs print the session total on exit.

How do I pick a model?

Put --model before the harness name; after it, flags belong to the agent. Auto Router is the default.

$ togetherlink --model zai-org/GLM-5.3 claude
What does it run on?

macOS and Linux. Everything lands in ~/.togetherlink/bin, with Bun brought along if you need it.

What is the daemon?

The local proxy that carries Claude Code and Claude Desktop traffic. It starts on demand and registers at login; daemon uninstall removes that.

$ togetherlink daemon status
How do updates work?

The binary keeps itself current on its own. One command forces it right now.

$ togetherlink update
What about telemetry?

Anonymous usage analytics tied to a random install ID. One environment variable turns it off completely.

$ export TOGETHERLINK_TELEMETRY_DISABLED=1

Your next session is one command away.

Free and MIT licensed · macOS and Linux · works with your Together API key