vibetrack MCP server
Commit Oct 13, 2026Lets an AI client query and monitor machine learning experiment results, including scalars, media, and system metrics.
Also known as chroneus/vibetrack
- Stars
- 29
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Oct 13, 2026
- License
- —
What is vibetrack MCP?
vibetrack lets an AI client query and monitor machine learning experiment results, including scalars, media, and system metrics.
Summary generated from the project's README
About vibetrack MCP server
Vibetrack provides a way for machine learning practitioners to track and monitor experiment results through an AI client. It allows users to query scalars, media, and system metrics such as CPU, GPU, and memory usage. The server can run locally while receiving data from remote training machines via a REST API, making it suitable for distributed training environments.
To use the MCP functionality, developers must install the package with the all extra and use Python 3.10 or higher. The server is launched using the vibetrack command with the mcp viewer flag. It uses SQLite and local files for storage, ensuring data remains in open formats. Users can also direct experiment outputs to various destinations like Slack, Telegram, or Jupyter using the SummaryWriter API.
What you can do with it
- 1Query experiment metricsAsk an AI agent to retrieve specific scalar values or system resource usage from a training run.
- 2Monitor hardware performanceHave an AI client check GPU utilization or memory consumption during a long training process.
- 3Analyze visual outputsRequest a comparison of image-to-image results generated during different experiment steps.
- 4Remote data ingestionSet up a central server to collect experiment logs from multiple remote training nodes.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachablechroneus/vibetrack is public and reachablegithub.com/chroneus/vibetrack ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)No licenseFinding
- Commit or release in the last yearLast commit or release 2026-10-13github.com/chroneus/vibetrack ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubvibetrack
Modern experiment tracking.

Key features:
- Send experiment results elsewhere: Telegram, Slack, Jupyter, Gradio, and MCP.
- Run locally while receiving experiment data over the network via REST API.
- Use open formats: experiment data is stored in SQLite and local files.
- Compare image-to-image results.
- Use a rich UI to show, hide, delete, and customize runs.
- TensorBoard SummaryWriter compatible drop-in APIs.
- Query results through the MCP server.
- Fast scalar logging; see the benchmark report.
Install
pip install vibetrack # default with web
pip install vibetrack[all] # all optional backends+dev; MCP on Python >=3.10
Quick start
TensorBoard-style API
from vibetrack import SummaryWriter
writer = SummaryWriter("runs/exp1", project_folder="my_project")
for step in range(100):
writer.add_scalar("loss", 1.0 / (step + 1), step)
writer.add_scalar("acc", step / 100, step)
writer.close()
See API.md for SummaryWriter and module-level logging examples.
Launch the dashboard
vibetrack --listen 0.0.0.0:6116
# -> Web UI on http://you_server:6116
# -> MCP is also mounted when installed with vibetrack[all] on Python 3.10+
Viewers and destinations syntax
writer = SummaryWriter("runs/exp1", project_folder="my_project")
writer.to("console").to("slack", every="15m")
writer.to("remote", url="http://server:8080", token="devtoken")
writer.add_scalar("loss", 0.5, step=0).to("telegram")
See VIEWERS.md for web, console, Slack, Telegram, Gradio, Jupyter, custom viewers, remote forwarding, credentials, and HTTP ingest.
vibetrack Architecture
flowchart TB
subgraph LOGGING["Experiment logging"]
direction TB
SW["SummaryWriter"]
ADD["add_(scalar|image|*)"]
SW --> ADD
end
subgraph REMOTE["remote HTTP logging"]
direction TB
TO_REMOTE[".to(#quot;remote#quot;)"]
end
VIBETRACK["vibetrack<br/>scalars • media • text"]
subgraph PERSIST["Persist"]
direction TB
DB_ROWS["scalars + text + metadata"]
DEFAULT_DB[("default DB<br/>~/.vibetrack/vibetrack.db")]
PROJECT_DB[("per project DB<br/>project_log_dir/vibetrack.db")]
MEDIA["media files<br/>project_log_dir/media"]
DB_ROWS --> DEFAULT_DB
DB_ROWS -.-> PROJECT_DB
end
DISPATCH[".to(viewer)"]
FANOUT["best-effort fanout"]
subgraph VIEWERS["live / summary viewers"]
direction TB
WEB["web<br/>(default)"]
CONSOLE["console"]
SLACK["slack"]
GRADIO["gradio"]
TELEGRAM["telegram"]
JUPYTER["jupyter"]
MCP["MCP"]
CUSTOM[".to(#quot;custom#quot;)"]
end
ADD --> VIBETRACK
TO_REMOTE -.-> VIBETRACK
VIBETRACK --> DB_ROWS
VIBETRACK --> MEDIA
ADD --> DISPATCH
DISPATCH --> FANOUT
FANOUT --> VIEWERS
DEFAULT_DB --> VIEWERS
PROJECT_DB -.-> VIEWERS
MEDIA --> VIEWERS
classDef logging fill:#fff4e6,stroke:#f59e0b,stroke-width:2px,color:#172033;
classDef code fill:#fff7ed,stroke:#f59e0b,stroke-width:1.5px,color:#172033,font-family:monospace;
classDef liveCode fill:#eef2ff,stroke:#6366f1,stroke-width:1.5px,color:#172033,font-family:monospace;
classDef event fill:#ffe4ec,stroke:#e11d48,stroke-width:2px,color:#172033;
classDef live fill:#eef2ff,stroke:#6366f1,stroke-width:2px,color:#172033;
classDef storage fill:#e7f8ef,stroke:#10b981,stroke-width:2px,color:#172033;
classDef optional fill:#f8fafc,stroke:#94a3b8,stroke-width:1.5px,stroke-dasharray:5 5,color:#64748b;
classDef optionalCode fill:#f8fafc,stroke:#94a3b8,stroke-width:1.5px,stroke-dasharray:5 5,color:#64748b,font-family:monospace;
classDef viewer fill:#f5e8ff,stroke:#a855f7,stroke-width:2px,color:#172033;
class SW,ADD code;
class TO_REMOTE optionalCode;
class VIBETRACK event;
class DISPATCH liveCode;
class FANOUT live;
class DB_ROWS,DEFAULT_DB,MEDIA storage;
class PROJECT_DB optional;
class WEB,CONSOLE viewer;
class SLACK,GRADIO,TELEGRAM,JUPYTER,MCP optional;
class CUSTOM optionalCode;
FAQ
How do I collect results from another server?
Run a vibetrack ingest server on the machine that should receive results:
vibetrack --listen 0.0.0.0:8080 --token devtoken --project-folder /srv/runs
Then forward events from the training machine:
from vibetrack import SummaryWriter
writer = SummaryWriter("runs/exp1", project_folder="my_project")
writer.to("remote", url="http://server:8080", token="devtoken")
See VIEWERS.md for remote forwarding and direct HTTP ingest.
How do I use separate databases on one machine?
By default, vibetrack writes to ~/.vibetrack/vibetrack.db. Pass
project_folder to keep a separate vibetrack.db inside that project folder.
from vibetrack import SummaryWriter
writer = SummaryWriter("runs/exp1", project_folder="projects/cv")
Open that database in the dashboard with the same folder:
vibetrack --project-folder projects/cv
How does vibetrack work with distributed training?
vibetrack automatically detects RANK / LOCAL_RANK. Only rank 0 logs by
default; other ranks get a silent no-op writer.
from vibetrack import SummaryWriter
writer = SummaryWriter("runs/distributed", project_folder="project/")
writer.add_scalar("loss", loss.item(), step)
writer.close()
To force every rank to log, pass rank="all":
writer = SummaryWriter("runs/distributed", rank="all")
Can vibetrack collect system resources?
Yes. vibetrack can collect CPU, GPU, memory, and project disk metrics in a background thread.
writer = SummaryWriter("runs/exp1", system_metrics_interval=3600) # every hour
Collected metrics include system/cpu_percent, system/mem_used_gb,
system/disk_free_gb, gpu/utilization, gpu/memory_used_gb, and
gpu/temperature. Use system_metrics_interval=0 to disable collection.
How do I run the MCP server for an LLM agent?
Install MCP dependencies, then run the standalone MCP server:
pip install vibetrack[all]
vibetrack --viewer mcp --project-folder my_project/
See MCP.md for MCP endpoints, tools, resources, and agent usage.
What CLI commands are available?
vibetrack # default
vibetrack [PROJECT_FOLDER] # Launch dashboard (web + ingest; MCP with vibetrack[all] on Python 3.10+)
vibetrack --port 8080 # Custom port
vibetrack --token SECRET # Protect ingest endpoints
vibetrack --listen 0.0.0.0:9009 # Open server on separate port
vibetrack migrate PROJECT_FOLDER # Merge legacy per-run DBs into project DB
How is vibetrack configured?
Settings are stored in ~/.vibetrack/config.json. The web UI can also write
project-scoped settings through its Settings tab.
{
"smoothing": "ema",
"smooth_weight": 0.6,
"system_metrics_interval": 3600,
"web": {
"theme": "light",
"auto_refresh": 5,
"image_play_fps": 2,
"original_values_opacity": 0.17
}
}
License
Apache 2.0
vibetrack MCP questions
Can vibetrack collect system resource metrics?
Yes, it can collect CPU, GPU, memory, and disk metrics in a background thread by setting the system_metrics_interval parameter.
How does vibetrack handle distributed training?
It automatically detects RANK and LOCAL_RANK, defaulting to only logging from rank 0, but you can force all ranks to log by passing rank="all".