Multiple Google accounts, kept separate

You can now work across several Google accounts in Redeep instead of collapsing them into one blurred connection. Connect a personal account and a work account — or two work accounts — and switch between them for Gmail and Calendar as the task requires.

Anything that changes the state of your accounts — sending mail, creating or moving an event — stays approval-gated. And accounts disconnect individually: removing the client's shared inbox leaves your own connection untouched.

The practical difference: “schedule this with the design team” and “reply from my personal address” can now be two different accounts in the same workspace, without signing out and back in between them.

Claude Code and Codex, working inside your workspace

External coding agents are no longer bolted on beside Redeep — they run through the same tool surface your other agents use. Claude Code and Codex subagents now have fuller MCP support, including host tool forwarding, so a delegated agent can reach the tools and context of the workspace it was launched from.

Credentials work the same way. Redeep can share the credentials and keys you have already stored, so a run doesn't require you to insecurely copy & paste keys into a terminal. Their work is tracked as activity you can follow on the card rather than a black box that returns a wall of text at the end.

Jobs that you run also notice earlier iterations on the same card, so a second pass builds on the first instead of re-seeding work that is already done. If you are new to delegating this way, the guide to running parallel agents covers the card-per-task pattern these agents plug into.

An interrupted turn does not have to start over

When a turn is stopped — by a timeout, or because you interrupted it — Redeep can now keep resumable session state instead of dropping the work on the floor. You pick the thread back up rather than re-explaining the task from the beginning.

The activity stream got the same treatment from the other direction. Redesigned activity cards surface the major thoughts and updates, so what an agent is actually doing stays visible without the transcript burying it. Less scrolling to answer “where is this up to?”

A few more improvements in 0.2.6

The supporting changes are smaller, but several of them will show up in everyday use.

  • Use tools in code support: more direct Python and data-analysis work, with built-in access to GitHub metadata, boards, cards, and persistent agent inspections.
  • New model options: Meta Muse Spark 1.2 with vision and agentic capabilities, plus updated Google Gemini 3.7 Flash routing.
  • Better routing and caching: provider ordering now accounts for eviction tracking and host affinity, with updated cache-hit scoring — lower latency and more efficient prompt caching.
  • Reliability work: steadier MCP server lifecycle and process management, more dependable merge gates, and batch file operations that handle duplicate creates and conflicts with structured overwrite safety.

Not every change lands in every workflow, and this is a summary rather than the full changelog. The complete list ships with the update in Redeep's release notes.

How to get Redeep 0.2.6

If you already run Redeep, the built-in updater will offer 0.2.6 — accept it and you are done. Your boards, sessions, and connected accounts carry over.

New here? Download the free macOS build. Redeep needs macOS 14 or later on Apple Silicon, and you bring your own provider key — or run a local model on-device with no key at all. The documentation covers first-run setup.