Qoopia releases

Version 5.0.25 — 11 November 2026

A reliability release: a general audit of the whole product and 84 fixes. Memory from Claude Code and Codex keeps arriving after a profile is linked to another agent, and qoopia doctor reports a session whose delivery stands still. qoopia_execute now runs in a separate process with a memory ceiling: a script that loops or takes too much memory is stopped, and the server keeps answering.

Long input no longer stalls file_put or the secret check, and ordinary names are no longer mistaken for Qoopia keys. Agents on a computer keep waking when one of them has lost its CLI or folder. Sign-in limits can no longer be used to lock an owner out. The dashboard keeps a key sign-in after the first page and does not redraw a page while you work in it. Old backup copies on other computers are removed. Database schema 56, instruction kit 18.

Version 5.0.24 — 7 November 2026

Agents get a much shorter tool list. A direct MCP client now sees eight core tools instead of the full catalog, plus qoopia_protocol for older connections. Every other tool still works by its old name, and tool_search finds the ones a task needs. Existing connections keep working without reconnecting.

qoopia_execute combines several searches and reads in one call inside an isolated sandbox that can only read, with the agent's own permissions. note_get reads up to 50 notes at once, and recall answers in well under a second unless the agent asks for the memory model's judgement. MCP clients can register without a key again (community issue #37); access still needs your consent. Retried messages are no longer stored twice. Database schema 56, instruction kit 17.

Version 5.0.23 — 24 October 2026

Work passes between Claude Code and Codex. When a session opens in a project, it receives the latest work of your other agent there: what was done, what is open, its final answer, your last requests, tool errors and the files it changed, each with a content digest checked against this computer. It is marked as history, not instructions. Clients without hooks get the same through work_handoff; work_take marks work as taken without locking anything.

Working notes are facts from the session record plus a model summary, and the model can no longer add a limit of its own. Secrets are removed before anything is stored. Very large transcripts are sent in small portions without slowing your agent; what does not fit goes on in the background, and doctor tells a server that does not answer from delivery that is only behind. Search treats ё and е alike; you can delete your account from the iPhone app. Database schema 55, instruction kit 16.

Version 5.0.22 — 18 October 2026

Canon is now a view-and-retire page. Ask your agent to write a rule; the night-review agent proposes drafts with reasons and sources, and only the owner accepts them in Overview. Owner sign-in renews when used, and secret rotation can keep existing devices signed in with a transition key. Access shows API-key issue dates without inventing dates for older keys.

Memory-model sign-in has expiry warnings where the provider exposes a date and a clear recovery message when sign-in is required. The dashboard has consistent English and Russian text and preserves your place when older messages load. Mac offers Open at Login once. Backups support optional destinations, bounded copy waits and consistent SQLite snapshots. The site uses a smaller WOFF2 font, correctly localized screenshots and real 404 responses. A nested JSON tool-result regression protects continuity summaries. Database schema 53, instruction kit 15.

Version 5.0.21 — 15 October 2026

Canon: short rules that Qoopia keeps in front of your agents by itself. A rule belongs to the whole workspace, one agent or one project, and arrives at the start of a session or on the turns it applies to. Agents write their own rules, and you edit them on the new Canon page. Hermes agents now save their conversations to memory on their own. A night-review agent, which only the owner can appoint, reads the workspace’s conversations; every other agent sees only its own.

A full audit brought fixes across the server, the clients, the dashboard and security: hooks stay fast when the server hangs, the wake-runner survives restarts and broken jobs, the dashboard speaks your language and gives agents and sessions their own addresses, and a long list of security checks is closed. The website counts installer downloads without cookies. Database schema 51, instruction kit 14.

Version 5.0.20 — 11 October 2026

Agents now wake each other: a message in AgentComm starts a real turn of Claude Code or Codex on your computer through the new qoopia wake-runner, and the agent answers by itself. Every wake is signed, only a request is answered, and delivery is reported honestly: woken, read from the inbox without a wake, or no runtime to wake. Every connection, including ChatGPT and Claude, can now send, read its inbox and reply.

A full audit brought about ninety fixes, each with its own test: a sign-in email can no longer enroll someone else's device, OAuth client registrations and the consent page are bounded, PDF and DOCX text is read in a separate limited process, memory status names the real cause, and the dashboard explains errors in plain words. The website has real Russian pages. Database schema 49, instruction kit 13.

Version 5.0.19 — 10 October 2026

Safer file updates: concurrent appends detect a conflict and ask you to retry instead of overwriting another update. File changes and their activity records are saved together.

Agent replies stay addressed to the caller’s conversation. Retried messages handle topic whitespace consistently. OAuth form validation and consent-response limits are stricter; the MCP client uses SDK 1.31 with the issuer fix. Database schema 48, instruction kit 12.

Version 5.0.18 — 6 October 2026

Autosave works out of the box with ChatGPT, Claude.ai, Grok bots and Muse: while an agent's Autosave switch is on, Qoopia asks its model to save every turn of the conversation, and one switch turns it off at once. The dashboard says "Saving automatically" only for agents that really save.

Claude.ai connects through a connection address with a trailing slash; Codex on the computer that runs Qoopia saves its sessions locally. A shared Linux host no longer lets another user sign in to your dashboard with a code you approve; knowledge pages change only by their author, the steward and you. Qoopia speaks MCP 2026-07-28.

Existing agents, connections, memory and permissions are retained. Instruction kit revision 12; database schema 48.

Version 5.0.17 — 4 October 2026

External access works end to end: ChatGPT and Claude.ai connect to a Mac or Linux install behind its tunnel, the consent page opens in your language and confirms from the browser you started in. Claude Code and Codex on the same computer save their sessions locally, and keep working after updates.

Logs and daily backups survive a Mac restart, sleep and Time Machine. A Linux server without a screen signs in with a code confirmed on your phone. Neighbouring agents can read each other's notes but only the author, the steward and you can change them; agent names may be Cyrillic; Russian search finds other word forms.

Security and reliability fixes across connections, restore and the phone dashboard. Existing agents, connections, memory and permissions are retained; Codex asks once to approve the Qoopia hooks again. Instruction kit revision 11; database schema 48.

Version 5.0.16 — 3 October 2026

Sign in with Google and you are in Qoopia straight away: no confirmation email, no code. Sign in once on each device; website and iPhone sessions now last a year and renew on every visit. A sign-in finishes only on the network where it started, so someone else's link signs nobody in.

A clearer dashboard: you first, then your steward and agents in a fixed order with memory switches right in the row; an overview with live activity; one list of connected apps with Disconnect; folders in Files, where agents can save files themselves.

Memory summarises only live conversations, so finished ones no longer spend your model subscription; recall is faster on Linux. Existing agents, connections, memory and permissions are retained. Instruction kit revision 9; database schema 48.

Version 5.0.15 — 2 October 2026

Shared context: every agent reads its workspace siblings' notes and conversations by default, so any agent can pick up another's work. The owner turns it off per agent on the agent page, or asks the steward; no agent reads another workspace.

A full audit of the product: security hardening of sign-in, sessions and connections, faster dashboard search and lists, accessibility fixes, clearer install and update messages, safer updates that refuse downgrades and keep a rolled-back version.

Existing agents, connections, memory and permissions are retained. Instruction kit revision 9; database schema 48 (applied by the update).

Version 5.0.14 — 30 September 2026

Connect an agent with one setup request. Qoopia confirms its first successful protocol call automatically; temporary verification codes are no longer part of dashboard setup.

Approve the new application's permissions using your existing signed-in owner session. Pending applications stay visible, status updates automatically, and failed clipboard access provides a manual copy option.

Instruction refresh preserves each agent's identity and container-local paths. Existing agents, connections, memory and permissions are retained. Instruction kit revision 7; database schema 47.

Version 5.0.13 — 30 September 2026

Muse.app and Muse Code CLI now have distinct setup paths. Give a connection the agent's name and preserve its identity when editing its label.

Your steward can prepare one owner review link. Cloud Muse and Grok Bot receive setup and verification in one request; active native connections and unused drafts are shown separately.

Memory checkpoints recover from timeouts using smaller batches. Existing data, permissions and verified connections are preserved; instruction kit revision 6, schema 47.

Version 5.0.12 — 24 September 2026

Connect Muse Code or Grok Bot through a guided, per-client MCP link. Existing connections and their authorizations are preserved.

My Qoopia agent responds more smoothly in the dashboard and Telegram: progress appears while it works, chat refreshes without waiting for a slow Send, and stalled native turns end with a clear recovery message. Files load when opened.

Profile and the owner-only panel open within your signed-in dashboard. iPhone TestFlight availability is tracked separately from Mac and Linux updates.

Version 5.0.11 — 23 September 2026

Profile now opens inside your signed-in dashboard on Mac, web and mobile. See your workspace address and sign out without another email sign-in.

The separate owner panel still checks your account permissions.

Version 5.0.10 — 21 September 2026

Connected Codex and Claude Code profiles now refresh their managed Qoopia instructions at session start after updating the app. Your instructions, agent roles and memory settings are preserved. Edited files are reported for review.

Clearer protocol version reporting helps identify outdated clients. Release metadata now verifies the database schema against both signed installers.

Version 5.0.9 — 21 September 2026

A consistent Graphite identity

Theme-aware favicons, the approved logo and consistent typography across the dashboard, account pages and new emails.

Also included: changes since 5.0.5

• Automatic memory resumes without dropping the first batch after manual mode. Manual-period content stays excluded.
• Mac upgrades from older signed installations open correctly and preserve existing data and connections.
• Overview opens first. Compact navigation and persistent chat keep your agent within reach on every dashboard page.
• Email confirmation opens the connected dashboard directly in the iPhone app. Native iOS uses a separate TestFlight build and Apple review.

Check iPhone availability →

Version 5.0.3 — 15 September 2026

Stable installers for Mac and Linux are public. Authorization checks are consolidated, and administrative write access is preserved consistently. Dashboard sessions and recall settings are separated, unused code is removed, and a TypeScript check of runtime dependencies rejects new cycles. Release metadata is generated only from authenticated package inventories with a matching source commit. Database schema 43.