Talyx
This server does not currently have a deployable release on Glama. Follow the repository's setup instructions to install and configure it for your MCP client, including any required credentials and local dependencies.
Talyx is a desktop app for a team of AI bots. Each bot has its own job, model and look. Talyx, the manager bot, hands work to the right one and briefs you when it is done. The bots run on an agent engine that the first-run setup installs for you.
If Talyx is useful to you, a ⭐ helps others find it.
What you get
A roster, not a chat list. Every bot is a block in the sidebar, with pins, folders, unread marks and what it is doing right now.
A manager bot. Ask Talyx. It asks the right bots and comes back with one brief.
Rooms. Put several bots in one room. Each answers in turn, and
@addresses one.You stay in charge. A bot asks before it runs a command, and asks you when it needs a decision. Routines run on a schedule.
Your providers. API keys, account sign-ins or your own endpoint. Pick a model per bot.
Agents can drive it. Claude Code, Cursor and others operate Talyx over MCP.
Local and open. No account, no server, no telemetry. AGPL-3.0.
Related MCP server: AKIMCP
Install
Builds come from the GitHub releases.
macOS 13 or newer. Open the DMG and drag Talyx to Applications, then open it. The Mac builds are signed with Developer ID, run with the hardened runtime and are notarized by Apple, so macOS opens them without a warning.
Linux (x64). Download
Talyx-x86_64.AppImage, make it executable (chmod +x) and run it, or installTalyx-amd64.debwithsudo apt install ./Talyx-amd64.deb. There is no Windows build yet: it comes once the engine runs there.Arch Linux: coming soon. The AUR package
talyx-binis written (packaging/aur) but not published yet; until it is, use the AppImage. It repackages the release's deb, pinned by its sha256. Like the deb, it does not update itself: update it with your AUR helper. There is no Flatpak: the engine runs from your home folder, which the Flatpak sandbox blocks (see packaging/aur).
A new release shows in the footer update button. Nothing downloads until you click Download, and Talyx never restarts by itself. The deb updates from the release page. Updates has how the check works and how each download is verified.
Engine
Talyx runs its bots on the engine, a separate agent backend that Talyx's setup screen installs for you. Talyx needs engine 0.22.7.
The engine's command must be reachable as
talyx-engine: on yourPATH, in~/.local/bin/talyx-engine, or picked with Choose file on the first-run screen. The setup screen installs the engine and links it astalyx-engine.Talyx attaches to the engine's gateway on this machine with
talyx-engine gateway ensure --json. The gateway runs on its own and outlives Talyx: quitting Talyx never stops it, and the next launch attaches to the same process.Talyx reads every path and port from the engine's own output; it never assumes where the engine keeps its data.
Windows cannot attach yet: minting a gateway ticket there goes through the engine's own runtime, so a Windows build comes once the engine runs there.
Agents operate Talyx (MCP)
While Talyx is open it runs an MCP server on http://127.0.0.1:47816/mcp (loopback only; TALYX_MCP_PORT changes the port). Its tools do what the app does: list the roster; create, rename and delete bots, folders and rooms; move, order and pin them; change a bot's title, description, SOUL, model, reasoning and look; list, add, pause and delete routines; rename a group room, change its bots, post in it and read its latest messages. Each tool runs through the same code as the roster and the details panel.
Your bots get the server in their profile automatically (Talyx adds it to every profile, also to bots created later), with its key kept in the profile's
.env.Outside agents (Claude Code, Cursor and others): Settings → Engine → Agents → Copy the add line, then run it in a terminal. Every request needs that key, and a request from a browser page is refused.
Approval: deleting a bot, a folder, a room or a routine, replacing a SOUL, a model the engine marks as costly, and adding or resuming a routine (it runs on its own, also with Talyx closed) ask you in the Talyx window first. Deny, closing the dialog, or leaving it for nine minutes changes nothing.
Talyx closed: the server is not running, so agents cannot connect and nothing changes. With Talyx running but its window closed, every tool answers that Talyx's window is not open.
A room's bots: the engine has no call to change who is in an existing room, so Talyx ships a small engine plugin (
resources/engine-plugin/talyx-rooms). Talyx writes it into the engine'spluginsfolder and enables it the first time a room's bots change. It changes the room in place, in the engine's own room store and in one transaction, the way the engine renames a room: the room keeps its id and history, a bot that joins reads the history and answers from the next message, and the log records who joined and who left. It refuses while a bot in the room is answering. The plugin uses the engine's internal room-store functions, so an engine update can break it: unit tests cover its install and call, and before an engine bump add and remove a bot in a test room against the new engine by hand.
Re-syncing the vendored client
Two things change when the engine moves on, and only one of them is updated from the app. The engine itself is: each Talyx release pins one engine version (EXPECTED_ENGINE_VERSION in src/shared/engine.ts, and engine-install.json on the release), and when the attached engine is older, the footer update button offers that version and installs it when you click Install the engine. The vendored client is not: when the engine's protocol changes, a developer redoes the steps below by hand and ships a new Talyx release.
src/renderer/src/shared/engine/ is the engine desktop app's protocol client and API wrappers, cut from tag v0.22.7 (its MIT licence is in that folder). src/main/infrastructure/engine/gateway.ts is its gateway attach and ticket code. When the engine moves to a new version:
Take the same files from the engine desktop app at the new tag:
src/api/{canonical-protocol,client,profiles,sessions,cron,models,config}.ts,src/types/(the engine types file, astypes.ts),src/lib/{gateway-rpc,session-owner-stamp}.ts,src/store/transcript-tail.ts, the shared package's{json-rpc-gateway,json-rpc-channel,gateway-events,gateway-contract.generated,reconnect-backoff,session-http-mutations}.ts, andelectron/local-gateway.tsforgateway.ts. Take the protocol tests the same way intotests/unit/engine/.Rename every word that carries the engine's name to
engine(types, functions, comments, messages), point imports at the local files, and runnpx prettier --writeon them.Diff against this folder and keep Talyx's changes, each marked
Talyx:in a comment: main opens the gateway socket (socket.ts,socketFactory); REST goes throughhttp.ts; wire strings and metadata keys come fromengineWire()insrc/shared/engine.tsat runtime; the remote-registry code paths are cut.Run
npm run typecheck,npm run lintandnpm test, then send "ping" from Settings → Engine → Debug console.
Before merging an engine bump, run TALYX_SMOKE_INSTALL_FILE=engine-install.json npm run test:smoke and paste its summary in the PR.
The smoke test also creates a scratch bot, sends it a message and, with a provider, gets a reply. The reply step needs an OpenAI-compatible endpoint: TALYX_SMOKE_PROVIDER_KEY (required), TALYX_SMOKE_PROVIDER_URL (default https://api.anthropic.com/v1) and TALYX_SMOKE_PROVIDER_MODEL (default claude-haiku-4-5). Without a key it is skipped, and the summary says so. The Release workflow passes the SMOKE_PROVIDER_KEY secret.
Build from source
Node 22.12 or newer (Node 24 matches the Electron runtime).
npm ci
npm run dev # Electron + Vite with hot reload
npm test # unit tests
npm run build:mac # DMGs in dist/; unsigned unless a Developer ID is set up
npm run build:linux # AppImage and deb in dist/ (x64)npm run dev downloads Electron the first time it runs.
Release
Releases are built by the Release workflow on GitHub Actions: Actions → Release → Run workflow. It builds the macOS DMGs and the Linux AppImage and deb, runs the real-engine smoke test on Linux, and creates a draft release v<version> with the installers, the update feed (latest-mac.yml, latest-linux.yml, the macOS zips and their blockmaps), engine-install.json and SHA256SUMS. The run also keeps the AUR package's PKGBUILD and .SRCINFO, bumped to the release's deb, as its aur artifact; the owner pushes them to the AUR after publishing (packaging/aur). Sign its update feeds on your own machine with npm run sign-release (the key is never on GitHub; this adds latest-*.yml.sig), review the draft, then publish it. Installed copies see a release only once it is published: GitHub does not show drafts to them.
The release body is the optional highlights input (lines like
- **Lead.** Text., shown by the update button) abovedocs/release-notes/<version>.mdand GitHub's generated notes. The form takes one line; for several, rungh workflow run release.yml -f highlights="$(cat release-highlights.md)".The workflow needs the secret
ENGINE_INSTALL_JSON, the base64 ofengine-install.json. Add it once, and again whenever the file changes:gh secret set ENGINE_INSTALL_JSON < <(base64 -i engine-install.json). Each job checks it againstENGINE_INSTALL_SHA256and fails when it is missing or does not match.
npm run release (scripts/release.sh) is the local path: macOS DMGs only, from your own machine, with the same install-file check.
Privacy
The free app has no server, no account and no telemetry. Without a module license key, the only network calls Talyx makes itself are three to GitHub: the update check; the update itself from the release when you click Download (through the signed update feed, on macOS and in the AppImage); and, at first-run setup, the release's engine-install.json, the steps that install the engine. With a module license key entered, Talyx also reaches the module server (modules.talyx.app) to list and download the modules that key unlocks. What your bots send to their model providers is between the engine and those providers; Talyx itself sends nothing anywhere else. Third-party components are listed in CREDITS.md.
Contributing
Issues and pull requests are welcome. See CONTRIBUTING.md for setup, checks and conventions, and SECURITY.md to report a security issue privately.
Website
website/ is talyx.app (Vite + React):
cd website && npm ci && npm run buildStatus
Talyx 1.1 is for daily use on macOS and Linux. It attaches to the engine on the same computer, which its first-run setup installs. You get a roster of bots with pins, folders and unread marks; a chat with each bot, with approval and question cards, images, files and dictation; bot details with routines, memories, skills and a look of its own; group rooms the engine hosts; search across a bot's chat and rooms; Usage, with what each bot and provider spent; keyboard shortcuts; and Settings for model providers, the engine, launch at login and the theme. Not yet: Windows, remote engines, the model's reasoning in the transcript, voice mode and other languages. The release notes have the details.
Before the pivot
Until version 0.11.0 Talyx was a file organizer and a hotkey assistant for macOS. That app is kept, unchanged, on the branch legacy/file-organizer and the tag v0.11.0-legacy. Build it from there if you want it; it gets no further changes.
License
Talyx is free software under the GNU Affero General Public License, version 3 only (AGPL-3.0-only): see LICENSE. Copyright (C) 2026 Petar Georgiev Stoev.
The licence does not cover:
The Talyx name and logo, including the logo and icon files in
resources/,build/andwebsite/public/. All rights reserved. A fork needs its own name and logo, and must not suggest that Talyx endorses it.The website's text and images: the copy in
website/src/and the images inwebsite/public/. All rights reserved. The website's code is under the licence.Talyx modules. They are separate, closed-source add-ons that are not in this repository and have their own licence.
Releases up to 1.0.5 were published under the MIT licence and stay available under it. Third-party credits are in CREDITS.md. To contribute, see CONTRIBUTING.md and the Contributor Licence Agreement.
The engine is Hermes Agent by Nous Research (MIT), used as it ships.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
This server cannot be deployed
Maintenance
Related MCP Connectors
Use your Mac, Windows or Linux computer from ChatGPT, Claude or Codex: files, commands, documents.
Work on your own computers from Claude or ChatGPT: run commands, edit and search files.
Talk to your public-facing AI from any MCP client — Claude, ChatGPT, Cursor, Cline, Windsurf.
Real-time chat hub for AI agents — Claude Code, Cursor, Cline, Codex over MCP or REST.
Related MCP Servers
- AlicenseBqualityDmaintenanceEnables Claude to interact with an Obsidian vault via the Local REST API plugin, supporting file operations, search, periodic notes, commands, and semantic search.223,713 npmMIT
- AlicenseNot gradedqualityAmaintenanceEnables web-based AI chat clients such as Claude, ChatGPT, and Grok to read, edit, and run things on your own machine — files, shell commands, search, Git, SQLite, browser automation, DevTools, background tasks, localhost APIs, clipboard, and Postman — through a single OAuth-gated endpoint. Access is governed by a strict command whitelist and folder-scoped permissions rather than unrestricted shell access.MIT
- AlicenseNot gradedqualityBmaintenanceEnables other agents to list, read, start, and steer Codex chats in the macOS ChatGPT desktop app, including sending follow-ups, choosing model/reasoning/speed, interrupting work, and answering approvals. It can run locally or behind bearer-token-protected tunnels for remote monitoring and control.56 npmMIT
- AlicenseNot gradedqualityBmaintenanceEnables Claude Desktop or Claude Code to ask ChatGPT and use shared task-board and file tools, with approvals handled through the desktop app or a command-line helper.GPL 3.0