Runners that fit the job
Several runners per project, each with its own image and size. Run them on Kubernetes or Cloudflare Sandbox, or connect your own machine. Hours are metered against your plan.

Fluxius runs Claude Code and Codex in shared, persistent sessions — with a team-lead agent, approvals, runners, integrations and usage control built in.
Web app · Telegram bot · desktop and mobile apps in development
A real recording of the Fluxius web app with demo data. Interface labels are translated into English for this page.
Works with the tools you already use
A session is a shared channel between people and agents. Pick a stop and watch the real interface.
Teammates see what the agent does as it happens. Presence, “need help”, a message queue and sessions that persist — so nobody works alone, and nothing is lost when a tab closes.
Run in approve-dangerous mode and risky actions pause until someone with write access says yes — from the web, Telegram or your phone.
Runners, tools, integrations, accounts, usage and access — so teams can ship with agents without building the plumbing.
Several runners per project, each with its own image and size. Run them on Kubernetes or Cloudflare Sandbox, or connect your own machine. Hours are metered against your plan.

VS Code, a terminal, live preview of your ports, generated artifacts and — when the runner has it — a secure desktop (VNC), all in a panel next to the chat.
A task, an error or a merge request from a connected service becomes a session with its full context. Agents read your tools directly — and write only after a human says yes.

Connect Anthropic, ChatGPT (device login), OpenRouter or Z.ai accounts per project. A run takes the first account that is under its limit and moves on when one runs out. Keys are encrypted and never shown again.

Tokens and cost by project, person, model and day — plus how much of each subscription window is left, and monthly budgets that stop runs at the limit.

Agents answer with cards like this one. Pin them to your home screen as widgets, then drag, resize and publish them to the project.

Give each project agents with their own avatar, prompt, models and skills — then choose one when you start a session, and keep a session private if it’s just for you.

Search, filter to “only mine” or by status, sort, go compact, pin favorites, reorder by drag. The master session always sits first, marked with a crown.

Scheduled tasks on a cron, Advisor mode for autonomous goals, and wake-ups so an agent can come back to a result later.
Keep an active session’s prompt cache warm so the next turn is faster and cheaper. Off by default.
Owners, developers and viewers per project, with Keycloak groups. Keep a session personal, then share it read or write.
Agents read your designs and comments without a plugin. With the Fluxius plugin open they can change the design too — and every change can be undone from the project’s Figma tab.
Talk to agents and answer their questions from Telegram. Building a Telegram Mini App? Test it inside the Preview panel with an emulated Telegram client and fake init data.
The interface comes in Russian and English, in light, dark or system theme with your accent color. Agents answer in the language you choose, in every session.

Use Fluxius from your own Claude Code, Codex or Cursor through an API token and an MCP endpoint — with your permissions, narrowed by scope.

Agents remember how you like to work — across projects and sessions. It’s yours alone: other members, admins and other people’s agents never see it, and you can export or erase it any time.
Browser notifications and Telegram when a run finishes or fails, when an agent needs approval or asks a question, and when a spending limit is close.
Connect repositories, add secrets and MCP servers, pick a runner and plug in Jira, Sentry and GitLab.
Pick models, write the system prompt, add skills and sub-agents, set permissions.
Blank, or straight from a ticket, an error or an MR. Invite teammates, or let the team lead start it.
Check the diff, approve risky steps, push and open a merge request.
Fluxius doesn’t replace your coding agent — it runs Claude Code and Codex for you inside a shared workspace and adds what a terminal on one laptop can’t.
Each session works in its own git worktree, so parallel tasks never step on each other.
Risky actions wait for a person with write access. Codex runs in a sandbox, and Fluxius asks before MCP tool calls.
Project secrets and provider credentials are stored encrypted and stay out of the frontend and the logs.
Who logged in, who changed an agent or a secret, who approved what, and every tool call the agents made.
Monthly token and cost limits per project, person or agent; owners are warned at 80% and runs stop at 100%.
Access follows your Keycloak groups. Personal sessions are invisible to everyone, admins included, until the author shares them.
Every session gets its own branch and working copy. Agents get no more access than the token you give them.
The web app is the product. Desktop (Tauri) and mobile (Capacitor) apps are thin shells around the same interface and are still in development.
A personal assistant that isn’t tied to a single project’s repository — for the questions that don’t belong to one codebase.
Bring Fluxius sessions into your editor, next to the code.
No. Fluxius runs Claude Code (via the Claude Agent SDK) and Codex (via the Codex SDK) for you in a shared workspace. You keep the agents; Fluxius adds sessions, collaboration, approvals, runners, integrations and usage control around them.
Two harnesses today: Claude Code and Codex, behind one runtime interface. Provider accounts can be Anthropic, OpenAI (including ChatGPT login), OpenRouter or Z.ai. Other terminal agents are not supported yet.
Yes. A session keeps its own history, so you can switch the model — including between Claude and Codex — and the next turn continues with full context. Both work on the same worktree, and every turn shows which model produced it.
The master session is a regular session whose tools are for managing other sessions: it lists them, starts new ones with a task, sends messages, interrupts and archives. Work sessions report back to it, and it can merge an MR under the project’s rules. It never gets more rights than your project settings allow.
That is configured per agent. In full-access mode it works autonomously. In approve-dangerous mode, Claude Code pauses risky actions until a teammate with write access approves. Codex has no per-command prompts in its SDK, so it runs inside a workspace-write sandbox and Fluxius asks for approval on MCP tool calls.
On runners: managed ones on Kubernetes or Cloudflare Sandbox, or your own machine. A project can have several, each with its own image and size; usage hours are metered against your plan.
Jira, Linear, YouTrack and Trello for tasks; GitLab, GitHub and Bitbucket for code and reviews; Sentry for errors; Grafana and Metabase for data; Confluence and Notion for docs; Figma for designs — plus any MCP server you add and a Telegram bot. A task, an error or an MR from a connected service can start a session with its full context. More services are on the catalog’s “soon” list.
In the browser today, plus a Telegram bot. Desktop (Windows, macOS) and mobile (iOS, Android) apps are in development as thin shells around the same web interface.
Sign-in goes through your company’s Keycloak. Access is by project, using groups for admins, owners, developers and viewers. Personal sessions are visible only to their author until shared.
Create a project, connect your repos and start your first shared session.