Changelog

MCP servers over HTTP

Zod now speaks the Model Context Protocol over plain HTTP — no local bridge daemon required.

• MCP-over-HTTP client — connect to any MCP server that exposes a JSON-RPC endpoint over HTTP (initialize, tools/list, tools/call).
• Settings → MCP tab — declare servers as YAML with a name, url, and optional headers.
• Session support — the client tracks Mcp-Session-Id headers and handles both JSON and SSE responses.
• Surface in the loop — each connected server's tools are offered to the model and executed on call, alongside built-ins and custom tools.

Workflow execution harness

Orchestration YAML no longer just edits — it runs. The runner turns multi-step harnesses into live agent workflows.

• YAML-to-steps runner — each step gets only its granted tools (schema-filtered), with a journal that carries context between steps.
• Control flow — @goto transitions with success, error, contains:, and match: conditions, plus a loop guard.
• Skill hints — steps can attach @skill entries that inject guidance into the step's system prompt.
• Live run card — an agentRun card streams step-by-step progress, per-step messages, and tool-run details into the chat.
• Launch points — Run from the orchestrator editor, the hub "Run Active Agent" quick action, a harness's play button, or a scheduled task.

Background service-worker agents

Scheduled tasks now keep running even when the tab is closed, powered by a classic service worker.

• Background scheduler — a service worker mirrors your localStorage state into IndexedDB and ticks due crons while you're away.
• Full agent loop offline — background runs execute tools and workflows against your configured provider, with up to 3 concurrent agents.
• Outbox sync — results land in an IndexedDB outbox; the page imports new sessions, messages, and cron status the next time you open it.
• Survives reload — a run that finishes while you're away appears as a done task with its reply when you come back.
• PWA sidekick — the same service worker keeps the app shell cached for offline use.

Custom tools & structured output

• Custom tool runtime — write JavaScript implementations in Settings → Tools; they execute in a sandbox with args, ctx state, fetch, and timers.
• Structured output cards — the model can emit [section] markers that render as thinking, findings, suggestions, to-dos, code blocks, files, errors, security warnings, and confirmations.
• Approve / Reject plans — a [confirmation] card shows a plan with a tool checklist you can approve or reject before anything runs.
• Aligned YAML editor — a hand-rolled formatter keeps every config block's colons aligned, wraps long values, and re-measures to the editor width.
• Socratic line editing — syntax-highlighted keys, masked secrets, clickable boolean toggles, and an "ask AI to edit this YAML" prompt on every section.