What changed and why

Desktop apps for macOS, Windows and Linux, fallback model, extensions from an archive

A Go core, ready-made apps and one-command install.

• One-command install: install.sh for Linux, macOS and WSL2, install.ps1 for Windows
• Apps: macOS (Apple Silicon and Intel), Windows, Linux (AppImage, deb, rpm)
• Fallback model: if the main one does not answer, the next one is used
• Stopping a task keeps the partial result
• Providers and models can be chosen with buttons in Telegram
• Extensions install from a ZIP or tar archive right in the chat
• Built-in connectors bundle

The wizard offers VPN+ when Telegram is blocked

If Telegram is unreachable during setup, the wizard offers to connect VPN+.

One address for the Telegram and model proxy

The proxy for Telegram and for the model is set with a single address.

Call recordings are transcribed out of the box

Call recordings are transcribed right after installation, with no extra setup. Our own Bot API server was added.

Files over 20 MB and sign-in with a ChatGPT subscription

Telegram accepts files over 20 MB. Sign-in with a ChatGPT subscription was added.

The Codex subscription stays connected

A maintenance release for defects found on live installations. The rotated Codex refresh token is now written back to auth.json: before, after a restart the agent presented a used-up token and the subscription was revoked with a 401.

• /connect is open to the owner and to people on the channel allowlist; the refusal names the owner's address
• The host's reply is recognised by the file timestamp, not by a change in content: reconnecting over an existing key works
• On timeout the agent distinguishes «the host did not reply» from «an older reply from <date> is on disk»
• The login watcher checks the request timestamp and stops on cancel; the request lands on disk atomically
• Long generations no longer break at three minutes: the blanket request ceiling is gone for every provider, and a deadline cut is not retried three times
• A provider that goes silent again hands over to the fallback: the answer comes from the fallback instead of a refusal half an hour later; the silent one is tried once, and its failure is visible in the cabinet and in doctor
• A provider that thinks for longer than a minute is no longer cut off: non-streaming providers now wait up to ten minutes for the first sign of a reply
• A button pressed by a person on a topic card runs immediately, without the model and without a second confirmation; topic deletion initiated by the model still asks
• Limitation: the button effect starts with new cards and only in Telegram; an already used-up token on older installations does not come back, a fresh login is needed

The panel, groomed: chat is the main screen

The conversation opens straight away, and the panel sections moved into the header and a slide-out drawer. Onboarding now happens inside the chat, and the agent remembers the answers instead of the browser: start on a laptop, continue on a phone.

• Files attach through the paperclip in the input field and keep the name you gave them
• The file list belongs to the current conversation, not to the whole agent
• Conversations live in a left-hand column, grouped by day and searchable
• System cards are compact: name, state and action on one line
• A system failure answers in plain words instead of a technical trace
• The English view translates the names and descriptions of built-in systems, the MCP catalog and the tool sets
• The terminal opens as its own full-screen view with a way back to the chat

Maintenance for access boundaries and pending actions

The accepted candidate closes Team profile and policy failures, persists pending approvals before exposing an action, and narrows the trusted desktop boundary.

• The HTTP guard and connector receive separate snapshots of exact arguments; full caller confirmation is still under review
• A child worker receives current authority but not the parent conversation or memory stores
• An unavailable pending-approval store returns HTTP 503 instead of a healthy empty list
• Privileged Electron IPC is available only to the trusted main document
• An explicitly selected credential file is not silently replaced by another source

A normal desktop app and a meaningful first launch

The release adds a complete desktop path for macOS, Windows and Linux: download a ready-made file, review a device inspection and build personal context from discovered material only after checking it.

• An Apple Silicon macOS DMG, Windows installer and Linux AppImage are published with the release
• An active local Codex sign-in through ChatGPT is selected automatically in desktop mode when no other provider is configured
• Rules and memory files appear as safe excerpts; secret-like files and lines are not read
• Discovered context first enters an editable draft and is saved with a separate action
• A configured agent opens chat immediately on later launches

The day that added more checks than capabilities

Web search, the browser and skills in conversation. But the main thing is elsewhere: over three days the same thing happened eleven times - code written, and nothing to invoke it.

• Web search and page fetching: Yandex Search API v2, SearXNG and Brave under one result shape
• Browser: reading script-rendered pages, and actions behind three locks
• An automated pre-release check looks for things declared and called by nobody and keeps them out of the build
• You confirm a specific action, not a tool type: «click «Checkout» on avito.ru», not «run browser__browser_click»
• Before clicking, the connector compares the label against the live page
• The extension no longer grants itself permissions

A service on a production server and a fallback model

The core runs as a service on a production server, scheduled jobs are executed by the agent, and a model outage no longer leaves you without an assistant.

• Scheduled jobs are executed by the agent, not by a stub
• A fallback model provider when the primary one fails

Installs as a service and survives the night

The release from which the assistant can be installed as a service and used daily.

• A channel crash now brings the process down: previously the channel died, the service showed as active, and the bot stayed silent for days
• Exits with a non-zero code, systemd restarts it
• The working loop is separated from connectors, which can be used independently of the core

The release about resilience

The assistant from v0.5.0 worked but did not survive load: one large connector response knocked out the dialogue, a long conversation lost its beginning, and the very first 429 reached the human as a refusal.

• A tool result is capped at a fraction of the model window, and truncation is marked honestly
• A long conversation is compacted rather than losing its beginning

CapyAgent became a personal assistant

The channel accepts free speech rather than only commands: it holds a conversation with the model, remembers it, sets reminders and calls connector tools.

• The agent loop: think, act, observe, with a mandatory ceiling on turns
• Model providers: an OpenAI-compatible adapter, MiMo, YandexGPT, GigaChat
• Memory, reminders and new channels
• A tool invented by the model comes back as the text «there is no such thing» rather than crashing the loop

Channels

A shared channel layer appeared, with Telegram as the first implementation.

• The sign of life is the time of the last successful exchange, not a «running» flag
• Connections go strictly over IPv4 through our own dialer inside the process: fixing /etc/gai.conf would fix the whole machine, and on a client's machine that is not acceptable

Local interface and a token

capyagent serve brings up HTTP strictly on 127.0.0.1. The address is nailed into the code rather than taken from configuration: the core holds keys to business CRMs.

• A token is mandatory and compared in constant time
• A local port without a token is reachable by any open tab
• CORS is allowed for extensions only

The Yandex.Direct connector

Four tools, written around four ways of losing data silently.

• 201 and 202 mean a queued report, not an error
• The error arrives in the response body with status 200
• API limits are counted before the request, not after a refusal

The first working slice

The core can bring up a connector and call its tools, and the Bitrix24 connector can discover a portal and export data without losing it silently.

• An MCP client over stdio: handshake, tools/list, tools/call
• CLI: version, connector tools, connector call
• Portal discovery before analysis