← All updates

Changelog

Claude Code and Codex in your workspace

Run Claude Code and Codex CLI sessions next to Prokop's own runtime, pick models for all three from one picker, and give every agent memory, skills, and learning out of the box.

v1.18.0

Claude Code sessions

Start a session on your installed claude CLI with the login it already has, then pick the model and effort in the session picker. Claude’s built-in tools, including its subagents, run through Prokop’s permission checks, and subagent activity is saved in linked sessions.

Claude Code sessions support image uploads, Stop, native /goal, manual compaction, fork, and Edit and Undo. Workspace memory, agent memory, session search, and agent skill management are available as tools, and the selected agent’s prompt and memory are added to Claude’s system prompt.

Codex CLI sessions

Start a session on your installed codex CLI (0.156.0 or newer). Command and file-change approvals go through Prokop, and Codex subagents are saved as linked, read-only timelines.

Codex sessions support image uploads, native goals with a token budget, manual compaction, fork, Edit, and Undo, with the same memory, session search, and skill tools as Claude Code.

One picker for every runtime

The model picker now has a rail for Prokop, Codex, and Claude. A session’s first message locks its runtime; model and effort stay changeable within it. Agents can pin a model from any runtime, and New Session follows the default agent’s pin.

Settings → Server → Harnesses shows each CLI’s installed version and lets you turn Codex or Claude off for new sessions and scheduled jobs without stopping existing sessions. Settings → Usage shows the 5-hour, weekly, and per-model limits of your Claude and Codex logins, with reset times.

Agents out of the box

Every primary agent now gets memory, skills, learning, and a home workspace automatically, with no promote step. Manage each agent’s capabilities, learning, and its USER.md and MEMORY.md in Settings → Agents. “Preconfigs” are called agents throughout the app.

Learning works on every runtime: a learner pinned to a Codex or Claude model runs its reviews there. Those reviews cannot be undone from learning history.

Permission modes

Standard, Extended, and Full replace the five-level auto-approve setting for sessions, workspaces, and scheduled jobs. Commands and file operations ask only when there is a concrete concern: a path outside the workspace, secrets, something destructive, or unknown code. The ask shows that concern and its evidence. Catastrophic commands such as rm -rf /* ask in every mode, and all three runtimes follow the same rules.

Also in this release

  • Branch history draws a commit rail with commits to push and to pull, branch chips, and a sync line such as “3 to push” or “Diverged” that opens Push and Pull.
  • Live updates for Git status, the file tree, scheduled jobs, session lists, and running indicators are pushed by the server to every open client.
  • Notifications and titles now cover Claude and Codex sessions too.
  • Fewer, merged tools: the built-in catalog drops from 24 to 18 tools, for example one edit tool instead of three.

Removed

  • The standalone OpenAI API and plain Z.AI providers. The catalog keeps DeepSeek, MiniMax, Z.AI (Coding Plan), and Codex.
  • Creating, renaming, or deleting providers in the app. Providers come from ~/.prokopai/models.json or a registry sync; model editing is unchanged.
  • The Workflow and Scheduling workspace capabilities. Scheduling jobs from the app is unchanged.

Update

Run this on the machine hosting your Prokop server:

prokop update

On first start, saved permission grants are cleared and existing settings map to the new modes. Claude Code sessions need Claude CLI 2.1.259 or newer and Codex sessions need Codex CLI 0.156.0 or newer, each already logged in on the server machine.