LOCAL RUNTIME FOR AGENTS

One runtime for all your coding agents.

Terminus runs on your machine and connects the agents, terminals and browsers you already use into a single orchestrated workspace.

free · local-first · open source

// scroll

// RUNTIME

A whole team of agents, working on one canvas.

This is the Terminus workspace, rebuilt node for node. Strong models take the heavy reasoning; lighter, cheaper models take the repetitive work. Every slide runs a complete workflow.

// AGENTS

It doesn't replace your coding agents. It orchestrates them.

Bring the agents you already run. Each one joins the runtime as a worker with a scoped set of permissions, and reports back with output you can inspect.

Codex
gpt-6-astra
Claude Code
opus-5-5
OpenCode
deepseek-33b
Gemini CLI
gemini-2.5-pro

any compatible tool joins the same way

// TOKEN USAGE

Every token, on every node, as it happens.

Terminus reads each agent's session and normalizes input, output, cache and reasoning across providers. Below is the real Usage screen, accumulating the four workflows above as one project.

LIVE FEED

each line is a node reporting usage
  • +1.39Kmobile.qaclaude-haiku-5-5in
  • +1.40Kmobile.qaclaude-haiku-5-5in
  • +1.39Kmobile.qaclaude-haiku-5-5in
  • +2.03Kmobile.qaclaude-haiku-5-5in
  • +2.24Kmobile.qaclaude-haiku-5-5in
  • +186mobile.qaclaude-haiku-5-5in
  • +3.05Kqa.leadclaude-sonnet-5-5cache
  • +6.09Kqa.leadclaude-sonnet-5-5cache

WHERE TO OPTIMIZE

  • 0147% of all tokens were served from cache — long-lived context is paying for itself.
  • 0250% went to heavy models. Anything repetitive in that share can move to a light model.
  • 03Light models did 10% of the work for a fraction of the context — that's the split you want.
  • 04Largest live context: frontend at 121K. Split or compact before it degrades.
Per nodeinput, output, cache read/write, reasoning and live context
Per providerClaude, Codex, OpenCode, Gemini… in one normalized audit
Always visiblethe project total sits in the status bar on every screen
Resetzero a project's counters before the next experiment

// CANVAS

Open the project once. It stays where you left it.

Terminals, browsers, agents and notes live side by side in space instead of stacked in tabs. The arrangement is saved to disk, so closing the window costs you nothing.

Terminalsreal shells, not emulated output
Browsersa live page the agent can drive
Agentseach in its own pane, running
Notescontext that stays with the project
state persists on disk

// LOCAL-FIRST

Nothing leaves your machine.

There is no Terminus backend. Your files, terminals, browsers and models stay where they are. External AI services are reached only through a key you supply yourself.

ON YOUR MACHINE

  • Project files and git history
  • Real terminals and browsers
  • Local agents and local models

OUTSIDE · OPTIONAL

  • External AI, via your own API key

Run your own multi-agent runtime.

Free. Local-first. Install it on your machine and point it at your project.

no account · no pricing · no subscription — Go to Feedback page