Mintokei — Your AI agent command center

/// Universal agent surface

Every coding CLI. One workspace.

Mintokei drives each CLI natively, so you switch tools, models and machines without leaving the chat. Bring your own CLI and your own keys.

Claude Code

Opus 4.8 (1M) · Sonnet 4.6 · Haiku 4.5

supported

Codex CLI

GPT-5.5 · GPT-5.3-Codex · o3 · o4-mini

supported

GitHub Copilot

GPT-5.x · Claude Opus 4.5–4.7

supported

Gemini CLI

Gemini 2.5 Pro · Flash

coming soon

OpenCode

Named agent profiles

supported

/// Local & remote runners

Run on your desktop. Or anywhere.

Every agent runs on a runner machine — your laptop, a beefy build box, or a cloud VM. Enroll a remote runner with a single token and it shows up in your machine picker, online, capacity and all.

One paste to enroll. Paste the enrollment token into the runner and start it — it connects back on its own, no inbound ports to open.
Per-machine capacity. Cap concurrent agents per machine; Mintokei schedules around it.
Survives disconnects. Reconnect and Mintokei reconciles in-flight agents — queued messages are redelivered.
Pick a machine per task. Heavy build? Send it to the big box. Quick edit? Keep it local.

settings · machines

Enroll a remote runner machine with a one-time token

/// Web · Desktop · Mobile

The same command center, on every screen.

A web app, a native desktop app, and an installable mobile app — all talking to the same self-hosted backend. Kick off a task at your desk, approve it from your phone.

mintokei · web in your browser

Web

Open it in any browser and point it at your self-hosted server — full live chat, streaming as the agent works.

payments-api · split view 2 live

/// Split view

Watch every session at once.

Stop tab-hopping between terminals. Drag chats into a split layout and watch two, three, four agents stream in parallel — even when they're running on different machines.

Drag-and-drop panes. Split, swap and resize; your layout persists per workspace.
Mixed agents. Claude Code on the left, Codex on the right, one workspace.
Live everywhere. Each pane streams its own tool calls, edits and reasoning.
One repo, many hands. Several agents collaborating on the same codebase.

/// Review changes

Approve every change. Line by line.

Agents move fast — Mintokei keeps you in control. Every edit lands as a reviewable diff, grouped by file, with the working tree or any commit as the base.

Inline or split diff. Switch the whole review between inline and side-by-side, with full syntax highlighting.
File tree + counts. See changed files and line counts at a glance across the whole working tree.
Approve risky tools. File writes and commands surface as permission prompts you accept or deny.
Multiple repos. Review changes across every repo in the workspace in one pass.

review changes · working tree · 3 files +25 −1

Reviewing an agent's diff across three files

orchestrator → mintokei-mcp connected

▸ create_agent_task("Research the OAuth migration", Copilot)

✓ started · laptop

▸ create_agent_task("Draft the migration plan", Claude Code · hetzner-gpu)

✓ started

▸ create_agent_task("Code-review the diff", Codex · build-box)

✓ started

▸ list_agent_task_messages(research)

→ "found 3 breaking changes — flagged in the plan"

Different CLIs. Different machines. One orchestrator.

▸

/// Plans & orchestration

Different CLIs. Different machines. One orchestrator.

Don't drive every step. Hand the wheel to an orchestrator agent — it spawns the right CLI for each job, runs them in parallel across your machines, and pulls their results back. All over MCP.

Right tool for each job. Spawn a researcher, a builder, a reviewer — each on the CLI that fits, all from one orchestrator.
It reads the results. The orchestrator pulls each agent's messages back, threads them together and decides what's next.
Driven over MCP. create_agent_task, list_agent_task_messages and more — every line in the terminal is a real MCP tool call you can make too.

/// And a lot more

A workspace that does the boring parts.

Preview the running app without a browser tab, fire tasks on a schedule or a webhook, and get pinged the moment an agent needs you.

localhost:8080 · live preview in-app browser

Live preview of a running web app, chat visible

Previewing an image (dashboard) in the stage

A scheduled webhook trigger before any run

The scheduled trigger after a run, showing run history

An agent pausing for a permission approval in-app

Approving the request from the Telegram chat

Live preview

Preview a running app inline — no browser tab — and open images, dashboards or video right in the stage. Even a dev server on a remote runner, tunnelled back to you.

/// Download