Runners, live previews and agent skills removed: Spedy is the layer your agents work in

We've removed the coding runners, the live previews and the agent skills from Spedy. The reason is the same one as for the AI chat a few months ago: you already have your agent setup. Claude Code in the terminal, Codex in CI, Cursor in the IDE — every team has made its choice, and none of them is waiting for Spedy to boot yet another agent in a container. What's missing isn't one more agent. What's missing is the place where its work lands: context in, pull request out, the hour on the right client.

That is exactly what Spedy is from now on. We don't run agents. We are the work and coordination layer for teams that work with their own AI agents.

What's gone

Everything that assumed an agent driven by Spedy:

• Coding runners: runner registration, runner tokens (rnt_), runner jobs with logs and stages, the docker-compose up setup, runner analytics and cost tracking.
• Runner teams and multi-agent pipelines: assigning an agent team to a ticket, trigger statuses, pipeline stages and the automatic re-runs after "changes requested".
• Live previews per ticket: preview environments, branch selection, DB snapshots, preview screenshots and videos on the PR, the preview agent with its permissions, hooks, and annotations on preview images.
• Epic orchestrator: epic runs that spread sub-tickets across a shared branch in waves.
• Agent skills: the skills library in the organization settings.
• External MCP servers and organization secrets: both were only read by runners. You no longer store LLM keys in Spedy.

The "AI agents" settings group (Runners & Agents, MCP servers, Agent skills) is gone with them, as are the "Runners" and "Preview" feature toggles per organization and project. On the website, the AI agent pipeline and preview pages have been removed.

Why

A runner operated by Spedy was always a compromise: a second agent setup next to the one your developers use anyway, with its own configuration, its own skills, its own secrets. Every improvement to Claude Code or Codex reached you in the terminal first — and the runner only once we had caught up.

And the bottleneck in agencies isn't typing. It's context, review and accountability: which of the eight clients is this? Which budget sits behind it? Who reviews the PR? Where does the hour show up at month end? Those are questions for a work platform, not for an agent — which is why that's where we focus.

What stays

Everything that gives your agents context and gives you the overview:

• The Spedy MCP server: tickets, boards, statuses, labels, milestones, custom fields, wiki, knowledge base, pull requests, pipelines, timers and work context — for Claude Code, Codex, Cursor, Claude Desktop or any other MCP client. Every action runs with the permissions of the user who connected the agent.
• Personal access tokens and OAuth: how your agent authenticates.
• Handoff to your coding tool: one click in the ticket header opens the ticket in your tool — with context, branch name and repo as the prompt. The agent reports back with a comment.
• Pull requests and pipelines: GitHub and GitLab PRs on the ticket, status moves along, CI runs in the pipelines dashboard — whether a human or an agent opened the PR.
• Automations: rules on ticket events whose result lands on the ticket.
• AI knowledge base: self-hosted embeddings, reachable by your agents via MCP.
• Time tracking with approval of agent time: what an agent books via MCP arrives as a draft in your approval queue — separate from your people's time, on the right client's board. It only counts once a human approves it.
• Project templates from your own template repo stay as well — just without the preview wording.

What you need to do

Nothing. The runner, preview and skills pages disappear automatically once the update is live. Tickets, comments, linked pull requests, time entries and knowledge entries are unchanged.

Three notes:

• Runner tokens no longer work. A running runner container gets an error on its next heartbeat. You can simply shut it down.
• Existing runner data is deleted — runners, jobs, teams, previews, snapshots, epic runs, skills, external MCP server configurations and organization secrets. If you want to keep runner logs, export them first.
• Connect agents via MCP: create a token under Account → API tokens and add it to your coding tool. Step-by-step guides for Claude Code, Cursor and Claude Desktop are in the documentation.

Less infrastructure, clearer roles: you bring the agent, Spedy brings context, budgets and the overview.