Everything the control plane does.
The whole surface, terminals and control first - all of it on your own machine, no account.
Run, watch, and steer every agent from one pane.
Not just a mirror. Launch the agent, watch it live, hand the keyboard around, and take it back - without leaving the dashboard.
One writer lease, handed across seats: a native shell, a dashboard tab, and a paired device.
Launch 27 AI CLI tools
Start Claude Code, Codex, Cursor, Gemini and 23 more as a real terminal straight from the dashboard. Or run observer in any native shell.
An observer run attaches to the daemon by default, so your native terminal stays interactive as the first seat.
Jump in - many seats, one writer
Every live daemon-owned terminal is joinable from any dashboard tab as an extra seat. Up to eight read-only viewers watch; exactly one writer drives at a time.
Tile up to nine live terminals in a drag-and-resize grid whose layout persists. Never concurrent editing.
Track the agent live
A draggable Session Cockpit floats over each AI-tool terminal, so the run's cost and context are visible while it is still running.
Live cost with an AI-vs-tool split, next-message cost band, context fill, a 5h/7d rate-limit gauge, cache-expiry countdowns, and process sparklines. The rate-limit gauge needs the proxy.
Files & Git, per terminal
Each terminal carries a read-only file tree and text viewer rooted at its project root, plus a Git panel for the same repo.
Branch, ahead/behind, file status, and 100 commits of history. Read-only in v1.
Native → local → remote
Hand the single writer lease across your native terminal, a local dashboard tab, and a paired remote device. Type in the native shell to reclaim it.
The seat you took from stays read-only, with take-back offered. A WSL daemon's terminals even render in a Windows browser.
Remote & mobile, on your tailnet
Reach the same dashboard - terminals included - from a paired device on your own tailnet. That is how you jump into a running session from a phone.
Remote is opt-in, tailnet-only, off by default (Tailscale serve, private HTTPS). Remote launch/control is a separate opt-in.
The number no other tool gets right.
Ten dashboard tabs, each built because someone running AI agents needed an answer they couldn't get any other way.
Not a guess - through the proxy, the provider's own returned token counts (cache, reasoning, long-context), costed at list price. Other tools estimate; documented accuracy tiers cover the rest.
Every dollar split three ways - input, cache read/write, and output - per model, per session.
Provider-reported, not guessed
We capture the provider's own reported token usage: the input/output split, cache reads vs writes, reasoning tokens, and long-context surcharges.
Through the proxy those are exact returned counts, not estimates. Every other tool carries a documented accuracy tier.
Costed at list price
Cost is derived from the captured tokens at published list price - per model, per agent, per project, per session.
When a provider changes how it bills, a migration re-computes history against the live wire.
Cache, read vs write
The cache tab attributes reads against writes and tells you when a warm prefix is about to expire.
Cache ratio, reads vs writes, prefix-survival, and keep-warm guidance.
Right-size the model. Gate the egress.
Not every turn needs your most expensive model, and not every command should reach the network. Both levers are flagged, not forced.
Govern your agents - observe by default, block when you enable it, reachable from a paired device on your own tailnet.
Model routing
SuperBased classifies each task and can preview a right-sized, same-shape model downshift before anything changes.
A counterfactual cost simulation runs first. Off by default; advises only once enabled, with same-shape enforcement opt-in.
Egress guardrails
Put a policy gate on risky commands, and on what a session's tool calls send over the network.
Guard/egress runs in observe mode by default; blocking is opt-in and channel-limited.
Watch your own agents - or an app you host.
The same binary covers two jobs. Turn on the one you need; both run locally by default.
Your developer tools
Point SuperBased at Claude Code, Cursor, Codex, Copilot and 29 more. Get provider-reported token usage costed at list price, session history, waste detection, and an MCP server your agent can query.
An LLM app you run
Route an admin- or org-hosted LLM app's traffic through SuperBased for OTLP trace ingest, an eval plane, and input-admission guardrails - budget, session limits, and content filtering.
36 supported tools. One hub.
Each integration is read in its tool's own native session format. Add a supported tool and SuperBased picks it up - no config, no proxy, no SDK rewrite.
Know exactly what you're measuring.
Compare
How SuperBased stacks up against ccusage, cccost, Langfuse, Helicone, and LiteLLM.
Compare tools →Glossary
Plain-English definitions: gross vs net input, cache read/write tokens, reasoning tokens, long-context surcharge.
Read the glossary →Quickstart
Install, run observer start, and open the dashboard in under five minutes.
Get started →The story of SuperBased, as a playable retro arcade - six worlds, twenty-four levels, six characters. A beloved corner of the site, kept exactly where it belongs.
► Play the arcade tour