MCP server de TaskHub
Commit 10 nov 2026Permite que los agentes de IA y los procesos locales informen su progreso, reciban comandos y muestren telemetría en vivo a través de una aplicación en la bandeja del sistema de Windows.
También conocido como FelixMUrphy/TaskHub
- Stars
- 0
- Herramientas
- —
- Descargas, 30 días
- —
- Última versión
- 4.2.0
- Último commit
- 10 nov 2026
- Licencia
- —
¿Qué es TaskHub MCP?
Permite que los agentes de IA y los procesos locales informen su progreso, reciban comandos y muestren telemetría en vivo a través de una aplicación en la bandeja del sistema de Windows.
Resumen generado a partir del README del proyecto
Acerca del MCP server de TaskHub
TaskHub ofrece una forma centralizada para que desarrolladores y agentes de IA monitoreen y controlen procesos en segundo plano en Windows. Funciona como una capa de gestión local que utiliza una pequeña API HTTP para recibir actualizaciones de progreso de entrenamientos, simulaciones, renders o compilaciones. Los usuarios pueden visualizar telemetría en vivo mediante un icono dedicado en la bandeja de Windows y un panel desplegable con gráficos, indicadores y medidores en tiempo real.
Esta herramienta está diseñada tanto para uso manual como para la integración con agentes de IA como Claude Code y Codex. Una vez configurado, los agentes pueden utilizar herramientas MCP específicas como task_start, notify y task_list para reportar su estado directamente al dashboard. El sistema opera totalmente de forma local vía loopback, por lo que no requiere cuentas en la nube, aunque puede conectarse a servicios externos como Telegram, Discord o GitHub Actions para notificaciones remotas. Es especialmente útil para gestionar tareas de larga duración que requieren intervención humana periódica o ajustes de parámetros mediante botones e inputs personalizados.
Qué puedes hacer con él
- 1Monitorear entrenamientos de IA de larga duraciónUn desarrollador puede rastrear el progreso del entrenamiento de machine learning con gráficos de pérdida en vivo y medidores de uso de VRAM.
- 2Interactuar con procesos mediante notificacionesUn usuario puede responder a una solicitud de tarea o cambiar un parámetro a través de una notificación de Windows o un mensaje de Telegram.
Escrito a partir del README del proyecto. Cada caso de uso se verifica con una cita del mismo.
Verificaciones automáticas
Ejecutado 15 nov 2026Informe completo →Aún no verificado. Un servidor necesita una puntuación de 80, ningún hallazgo grave y suficientes datos para que podamos revisarlo. Cómo funcionan las verificaciones · Qué corregir para obtener más puntos
- Repositorio de origen vinculado y accesiblefelixmurphy/taskhub is public and reachablegithub.com/felixmurphy/taskhub ↗Aprobado
- El paquete indica este repositorio como su origenNo verificado
- Procedencia de la compilación (atestación de Sigstore)No verificado
- Licencia de código abierto (OSI)No licenseHallazgo
- Commit o lanzamiento en el último añoLast commit or release 2026-11-10github.com/felixmurphy/taskhub ↗Aprobado
- OpenSSF ScorecardNo verificado
- El nombre del paquete no es similar a un paquete popularNo verificado
- No hay avisos críticos en el paquete ni en sus dependenciasNo verificado
- Sin instrucciones ocultas para el modelo0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersAprobado
- Las herramientas destructivas están etiquetadasNo verificado
- El endpoint remoto responde al handshake de MCPNo verificado
README
Ver en GitHubTaskHub
One place to watch and control everything that runs in the background on your Windows PC — training runs, simulations, renders, builds, backups, and the scripts your AI agents (Claude, Codex, ChatGPT) start for you.
Any process reports progress over a tiny local HTTP API and gets a tray icon, Windows notifications and a pop-up panel with live charts, gauges and its own buttons (pause, change a parameter, save a snapshot, kill). No cloud, no account, loopback only.
⬇ Download the latest setup.exe · What's new · API reference · Privacy · ❤ Sponsor
AI agent reading this? Claude Code / Codex: use the TaskHub MCP tools (
task_start,notify, …). If you don't see them, don't usecurl- ask the user to enable TaskHub tools (MCP) (panel → ⚙ Settings → AI agents) and to start a new session. Details: API.md → AI agents.
Why
- You start a 6-hour job and want to know when it finishes, fails or hangs — without watching a console.
- An AI agent launches long scripts for you; TaskHub gives them one documented way to report status and to receive your commands back ("lower the learning rate", "stop after this epoch").
- Several things run in parallel and you want one glance at all of them.
Features
- Tray icon with a progress ring and the number of active tasks; colour shows running / stalled / failed / done.
- Panel (click the tray icon): every task with progress, ETA, status, log and widgets — line/area/bar charts, gauges, stats with deltas, LEDs, key-value lists, tables, text.
- Controls declared by the task itself: buttons, number/text/select inputs, confirmations, shell commands, open-file/folder, graceful stop, pause/resume, kill. Values go back to the process over long-polling.
- Task links open in the browser; a file / folder link is shown in Explorer (never run), its address is
in the link's tooltip, and a shell command gets a typed value as plain text (no extra commands from
&,|…). - Windows notifications on finish / failure / stall by default; start, milestone (plain) and log notifications on request; the task's buttons and input fields in notifications if it asks; a live progress notification; quiet while the panel is open.
- Claude Code and Codex in one click: TaskHub tools for the agent (MCP) and "finished / needs your permission" notifications (hooks) — see below.
- Plan limits (off until you turn it on): how much of the 5-hour and weekly limits Claude Code and Codex have used, in the panel, with warnings from 75% and 90% — see below.
- Telegram: the same notifications on your phone through your own bot, with the task's buttons; a live progress
bar per task and its charts when it finishes;
/tasks,/mute(optionally only when you are away from the PC) — see Telegram. - Discord (off until you connect it): the same notifications and buttons as direct messages from your own bot — see Discord.
- GitHub Actions: the workflow runs of your repositories as tasks — progress and ETA from the usual duration of earlier runs (by jobs / steps when there are none), the failed step and the end of its log, cancel / re-run buttons, artifact and release links — see GitHub Actions.
- Runpod: your GPU pods as tasks — GPU, region, image, $/h, spent, uptime, SSH and Jupyter links, load charts, GPU idle warning with a Stop button, balance and runway with a low-balance warning, stop / start / restart (delete only in the panel, after typing the pod's name) — see Runpod.
- Vast.ai: your instances as tasks — the same as Runpod (GPU, location, image, $/h, spent, uptime, SSH, load charts with disk, GPU idle, balance / credit and runway, stop / start / reboot, delete only in the panel) — see Vast.ai.
- SSH (off until you turn it on; tested on a real OpenSSH server in WSL, not yet on a remote one): your own Linux hosts as tasks, read-only and with nothing installed on them: load, disks, GPU, tmux sessions; a notification when a job ends, a disk fills up or the connection is lost. TaskHub keeps no key. SSH.
- AWS (off until you turn it on; built from AWS's documentation, not yet checked on a real account): EC2 instances as tasks with a Forgotten? reminder; SageMaker training jobs, AWS Batch jobs, this month's cost and forecast, and AWS Budgets, each behind its own switch. Read-only. AWS.
- Balance guard (off until you turn it on): for Runpod and Vast.ai — a warning at your own balance threshold (balance, $/h, time left), and at the stop threshold the running machines are stopped (never deleted) after a notification with Stop now / Snooze (Windows, and Telegram / Discord when linked) — see Balance guard.
- Work hours (off until you turn it on): Start / Pause / Finish above the task list, week and month bars against your norm, overtime and the amount earned; a calendar to correct any day, month closing, public holidays you confirm, amounts in another currency, and reminders for a session left open — see Work hours.
- Stall detection (heartbeat timeout) and lost state when the process dies.
- Agent-friendly: an MCP server (
taskhub mcp), a nativetaskhubCLI on PATH, a stdlib-only Python client, plain HTTP for everything else;taskhub apiprints the docs an agent needs. - Update check: a notification when a new version is released (once a day; switch and Check now in Settings → System; Check now answers right there, also when you already have the latest version).
- Panel like the Windows notification center: docked to the right edge of the screen at full height, slides in and out (instant if Windows animations are off), width by dragging its left edge, a pin to keep it open (the pinned panel keeps its screen edge: maximized windows make room, other windows too if you switch that on in ⚙); blurred see-through background and cards (Panel transparency; Card transparency — in ⚙), light / dark theme and accent colour follow Windows (or set in ⚙). The settings say when Windows itself draws no blur (energy saver on, or Transparency effects off in Windows settings), with a link to that setting.
- Settings: ⚙ in the panel (or Settings in the tray menu) opens them next to it, one icon tab per section — appearance, notifications, messengers (Telegram, Discord), work hours, connectors, AI agents, system (autostart, updates, data folder, API reference), about; the tray menu keeps only the quick actions (open the panel, mute, settings, support, exit) in the Windows 11 style. While the panel is pinned, only their own button closes them. On battery the panel and settings appear and go at once, without the slide.
- 10 UI languages (English, Українська, Español, Deutsch, Français, Italiano, Português, Polski, 中文, 日本語).
- Per-user install, no admin rights, no .NET runtime needed; self-healing autostart.
Quick start
- Run
TaskHub-<version>-setup.exe→ Install (choose a folder, autostart, PATH and the Claude Code / Codex options). The setup window stays open with the progress and ends with Open TaskHub; uninstalling works the same way (with an optional Also delete my data). - Report something:
taskhub notify "Render" "finished in 42 min"
taskhub run --name "Build" -- cmake --build build -j 16 # wraps any command, parses % from its output
taskhub run --name "Shot 12" -- blender -b shot12.blend -s 1 -e 240 -a # frame, samples, overall %, s/frame
taskhub run knows Blender's background-render output by itself (no add-on in Blender): the frame
and its samples, overall progress when the frame range is known (-s/-e or -f), a seconds-per-frame chart and the
last saved file, render memory. Blender exits with 0 even when a render fails, so its Error: lines mark the task failed
(and taskhub run exits with 1; not the memory-leak note Blender prints at exit). Video output (FFmpeg) and stereo views
saved as separate files count by frame. Stop after this frame (panel / notification) ends Blender once the
frame is saved and the next one begins; after such a stop or a crash the task says where to resume and taskhub run prints the same command with -s
set to the next frame. --preset none turns this off.
from taskhub_client import Task, W, C # stdlib only; installed next to taskhub.exe
with Task("Train", total=100, unit="epochs", source="codex",
widgets=[W.chart("loss", "Loss", series=["train", "val"]), W.gauge("vram", "VRAM", max=24, unit="GB")],
controls=[C.signal("lr", "Set LR", input=C.number(default=1e-3))], pausable=True) as t:
t.on("lr", lambda v: set_lr(v))
for e in range(100):
t.wait_if_paused()
if t.stop_requested: break
t.progress(e + 1, status=f"epoch {e+1}", widgets={"loss": {"train": tl, "val": vl}, "vram": mem})
curl -s -X POST localhost:47800/api/notify -H "X-TaskHub: 1" -H "Authorization: Bearer $(taskhub token)" -d '{"title":"Done","message":"backup ok"}'
With AI agents — Claude Code and Codex: enable TaskHub tools (MCP) (next section), nothing else to do. Other
agents (ChatGPT, scripts, any tool with a shell) — add one line to AGENTS.md / custom instructions:
For any job longer than a minute, report it to TaskHub (
taskhub apiprints the docs; wrap commands withtaskhub run --name ... -- <cmd>).
Claude Code and Codex: one click
The setup dialog (and later Settings → AI agents) has independent options per agent. If an agent is used on your PC and you haven't chosen yet, the panel shows a card until you choose, one notification says so, and nothing is written until you press Connect:
| Option | What you get | Where TaskHub writes |
|---|---|---|
| Notifications (hooks) | a toast when the agent finishes a response or needs your permission | Claude Code: ~/.claude/settings.json · Codex: ~/.codex/hooks.json |
| Plan limits (off until you turn it on in Settings → AI agents, for each agent) | how much of the 5-hour and weekly limits is used, in the panel, yellow with a warning from 75%, red with a warning from 90% | Claude Code: ~/.claude/settings.json (statusLine) · Codex: nothing, TaskHub asks it directly |
| TaskHub tools (MCP) | the agent gets tools notify, task_start/update/finish, task_commands, task_wait, task_list and shows its long jobs in the panel by itself |
Claude Code: ~/.claude.json · Codex: ~/.codex/config.toml |
TaskHub adds — and on uninstall removes (Store edition: see below) — only its own entries and keeps everything else (backup of the original:
*.taskhub.bak, and the version from just before TaskHub's last write: *.taskhub.last). An existing empty settings file is left alone and reported as an error. What to do in the agent afterwards:
- Claude Code — start a new local session (desktop app → Code tab → environment Local, the CLI or VS Code).
/mcpshould listtaskhubas connected. Cloud sessions run on Anthropic's servers and can't reach TaskHub on your PC. - Plan limits — needs a Pro, Max or Ultra plan; the numbers appear after Claude Code's next reply and update only while a session runs, so an old reading says how old it is.
Your own status line keeps working unchanged. If you have none, Claude Code shows an empty one and hides hints such as
esc to interrupt. Details: API.md. - Plan limits in Codex — TaskHub asks Codex for the numbers every 10 minutes while Codex is running, with no model requests. This relies on a part of Codex that its makers call experimental, so a Codex update can stop it: the numbers then just get old, and everything else keeps working. With no numbers to show, the panel says why (not found on
PATH, not signed in, couldn't read). Details: API.md. - Codex — restart Codex. Hooks: Codex → Settings → Hooks → Trust all (Codex asks once for new hooks).
MCP: the
taskhubserver should appear among Codex's MCP servers.
Scripted: TaskHubHost.exe --integrations claude=on codex=on claude-mcp=on codex-mcp=off. Manual setup:
API.md → Claude Code hooks, Codex hooks, MCP server.
Telegram
Settings → Messengers → Telegram: create a bot with @BotFather (/newbot), paste its token, press
Connect, scan the QR code with your phone (or open the link on this PC) and press Start (in a private chat with the bot:
groups can't be linked, since every member could press the tasks' buttons). Everything is set up right on the settings
page. To link another account later (e.g. the one on your phone), press
New link; Disconnect removes the bot token from this PC. From then on every notification TaskHub shows also arrives in
that chat (only that chat can talk to the bot); a task's notification buttons become inline buttons — including
select options and yes/no; number / text inputs: the bot asks for the value and your reply goes to the task — and presses
go back to the task. A task with progress gets one silent message with a progress bar, ETA, status and the task's
buttons that is edited in place (no new message per update); when the task finishes, its charts arrive
as one picture (option Progress and charts, telegram_progress). Commands: /tasks, /mute, /unmute. Option: send only
when you are away from the PC (no keyboard / mouse for 5 min, telegram_away_minutes in config.json). The bot token
is kept in the Windows Credential Manager (if the Credential Manager is unavailable it stays in
config.json, with a warning in Settings); messages go from your PC straight to Telegram's Bot API.
TaskHub keeps to Telegram's pace (about one message a second, and as long as Telegram asks after Too many
requests) instead of dropping messages, and long texts are shortened so they always arrive.
Discord
Built from Discord's documentation, tested against a simulator and checked with a real bot.
Settings → Messengers → Discord: create an application with a bot in the Developer Portal, paste the bot's token, press Connect, add the bot to a server of yours (a bot can't write to someone it shares no server with), press Check again and then the Connect button in the bot's direct message. Only the application's owner can be linked, and only that person's presses in that chat count. Disconnect removes the token from this PC.
From then on every notification also arrives there. A task's notification buttons become buttons, a list becomes a list (up to 25 options), number and text inputs open Discord's input window, and what doesn't fit in 5 rows stays in the panel. A task with progress gets one silent message that is edited in place, and its charts arrive as a picture when it finishes (option Progress and charts, discord_progress). Commands: /tasks, /mute, /unmute. Option: send only when you are away from the PC (discord_away_only, discord_away_minutes).
The token is kept only in the Windows Credential Manager; without it TaskHub doesn't connect. Your PC talks straight to Discord: no public address, no server of ours. How to get a token: docs/keys/discord.md (step by step, revoking; українською).
GitHub Actions
Off until you turn it on. Settings → Connectors → GitHub Actions (everything on the settings page): list the repositories (owner/repo) and give access — your own
fine-grained token limited to those repositories (Actions: Read and
write, Contents: Read-only, Metadata: Read-only; kept in the Windows Credential Manager, never in config.json), or the token
the Git Credential Manager already has (broad rights: all your repositories). Your workflows stay unchanged. Each run becomes a
task with the repository, workflow, branch / tag, commit and actor; failures always notify (Windows, and Telegram / Discord when linked), successes
if you want. TaskHub talks only to api.github.com; the optional log tail is fetched from GitHub's log storage through the
one-time link GitHub gives, without the token. A notification warns a week before the token expires. A run waiting for
approval (e.g. from a fork) shows Waiting for approval on GitHub and is followed once approved; a renamed repository is
followed and reported (repository moved to …); GitHub's rate limits pause polling. Details:
API.md → Connectors.
Pull requests, read-only (two switches on the same page, both off by default): pull requests
waiting for your review (any repository the token can see; a fine-grained token sees only its own repositories and
public ones) and open Dependabot pull requests in the listed repositories (the token also needs Pull requests:
Read-only). Each is a task with a link and a
notification when it appears; it ends quietly when it leaves the list (Merged, Closed without merging, No longer
waiting for your review). Nothing is merged, approved or commented on.
How to get a token: docs/keys/github.md (step by step, minimum permissions, revoking; українською).
Runpod
Off until you turn it on. Settings → Connectors → Runpod: paste an API key from the
Runpod console (Settings → API Keys; kept in the Windows Credential Manager, never in
config.json; turning the connector off keeps it, Remove deletes it). A read-only key is enough to watch: TaskHub checks once (and daily) whether the key
may run actions and shows the Stop / Start / Restart / Delete buttons only if it may. Each pod is a task (rp-<pod id>):
active while it runs or starts, finished stopped when it stops and terminated when it is deleted; GPU (type × count),
region, image, price per hour, spent so far (≈, from Runpod's billing), uptime, the SSH command and https links to its HTTP
ports (Jupyter on 8888) through Runpod's proxy, and GPU / VRAM and CPU / RAM charts from TaskHub's own polls (every 30 s).
When every GPU of a pod stays under 5 % for 30 minutes (idle_minutes, idle_util) TaskHub warns GPU idle — on Windows and
in Telegram, with a Stop button. Balance and runway (balance ÷ current spend per hour) are shown in Settings and on the pods;
below 12 hours (low_runway_hours, or under low_balance $) TaskHub warns once — at $0 Runpod stops pods and terminates the
ones without a network volume. Stop, start and restart ask for a confirmation (stop and restart also work from the
notification, Telegram and Discord); delete works only in the panel, after typing the pod's name (never through the HTTP API). TaskHub talks only to
api.runpod.io: REST v2 for pods and billing, and GraphQL for the balance (graphql_balance, on by default; turn it off to
skip the balance). Details: API.md → Connectors.
How to get a key: docs/keys/runpod.md (step by step, minimum permissions, revoking; українською).
Vast.ai
Off until you turn it on. Settings → Connectors → Vast.ai: paste an API key from the Vast.ai console
(its API keys page; a key with instance_read, user_read and billing_read is enough to watch, actions need
instance_write). It is kept in the Windows Credential Manager, never in config.json; turning the connector off keeps it, Remove deletes it. Like Runpod,
TaskHub checks once (and daily) whether the key may run actions and shows the Stop / Start / Reboot / Delete buttons only if it
may. Each instance is a task (va-<id>, named by its label): active while it runs or starts, finished stopped when it stops
and terminated when it is destroyed; GPU (type × count), location, image, price per hour, spent so far (≈, from Vast.ai's
charges of the last 30 days), uptime, the SSH command, a link to the Vast.ai console, and GPU / VRAM and CPU / RAM / disk
charts from TaskHub's own polls (Vast.ai often reports no load — then the charts stay empty). GPU idle and the low-balance
warning work as for Runpod. Vast.ai reports a balance and a credit; TaskHub shows both and counts the credit for the
runway — at $0 Vast.ai stops instances (stopped ones still pay for storage), and without a payment method they are scheduled
for deletion. A start can wait in Scheduling when the host rented the GPU out meanwhile; TaskHub says so. TaskHub talks only
to console.vast.ai. Details: API.md → Connectors.
How to get a key: docs/keys/vastai.md (step by step, minimum permissions, revoking; українською).
AWS
Off until you turn it on. Built from AWS's documentation and tested against a simulator, not yet on a real account. Settings → Connectors → AWS: paste an access key of a dedicated IAM user with a read-only policy (how to get it). TaskHub starts, stops and changes nothing in your account. EC2 instances become tasks with their state and how long they have run: in every Region enabled in the account, or only those you list. One running longer than your limit (72 hours by default, 0 = off) gets a Forgotten? reminder, with a button for the ones meant to run. Each of these has its own switch and its own permissions: SageMaker training jobs (phase and time; SageMaker reports no percentage for ordinary training, so TaskHub shows none), AWS Batch jobs (real progress for array jobs), this month's cost and forecast from Cost Explorer (AWS charges $0.01 per request; TaskHub makes 2 at most every 8 hours, about $1.80 a month, and a restart doesn't pay again) and AWS Budgets (one notification when a budget goes over its limit). Details: API.md → Connectors. How to get a key: docs/keys/aws.md (step by step, minimum permissions, revoking; українською).
Preguntas sobre TaskHub MCP
¿Cómo activo TaskHub para Claude Code o Codex?
Puedes activar las herramientas de TaskHub (MCP) yendo al panel de TaskHub, seleccionando Settings, luego AI agents y conectando el agente específico.
¿TaskHub requiere conexión a internet o una cuenta?
No, funciona localmente mediante una API HTTP de loopback sin necesidad de servicios en la nube ni cuentas.