An open-source MCP server that gives R&D teams long-term context and insights on their time-series data.
接入你的智能体
從發布者安裝。不是 PlugMCP 端點。
遠端 URL 還是本機指令? 兩種安裝方式的差異
來自發布者
An open-source MCP server that gives R&D teams long-term context and insights on their time-series data.
Long-term context and insights on your time-series data: an MCP server that lets your AI agent query all your boat, robot, drone and vehicle logs with SQL, on your machine. NMEA 0183/2000, ROS 2, MCAP, PX4, ArduPilot, MAVLink, CAN with DBC, MDF4.
[!WARNING] Early alpha. Commands and on-disk formats may change. Issues welcome.
Connects your AI agent to your time-series data (telemetry, logs, measurements…) on your machine: ROS 2 bags (MCAP or .db3), MCAP (ROS 2, protobuf, JSON), PX4 ULog, MAVLink .tlog, ArduPilot DataFlash, Vector CAN logs (ASC, BLF) with your DBC files, NMEA 0183 and NMEA 2000, MDF4, CSV, JSON Lines, Parquet and Arrow IPC.
Use the agent you already work with (Claude Code, Codex, Cursor…), with its model and your subscription. It finds abnormal and similar recordings by comparing each one with its peers.
Everything your agent reads fills its context window, and a fuller window makes it slower, costlier and less accurate. So your agent's SQL runs on your machine and only the answer comes back, while full-resolution signals go straight to your viewer.
Written in Rust, on Apache DataFusion and Apache Parquet. Each recording is decoded once into Parquet files, and your agent queries them with SQL through DataFusion: answers across all your recordings, in milliseconds.
Your agent opens the original recording at the right time in the viewer you already use, through the viewer's MCP server (Rerun, Foxglove…). It also sends any query result straight to Rerun, at full resolution.
Tell your agent what the files don't say, for one recording or a whole folder: the machine, what changed, the conditions, the purpose of a test. Wise Bucket keeps it for every later session, and your agent compares recordings of the same machine.
flowchart TB
rec[/"Your recordings<br/>ROS 2, MCAP, ULog, CSV…"/] -->|"decoded once"| store
subgraph wb["Wise Bucket"]
store[("Parquet files")] --> engine["DataFusion<br/>SQL engine"] --> mcp["MCP server"]
end
you(["You"]) -.->|"CLI configuration"| wb
mcp -->|"SQL, in milliseconds"| agent["Your AI agent<br/>Claude Code, Codex, Cursor…"]
mcp -->|"query results"| viewer["Viewer<br/>Rerun, Foxglove"]
agent -->|"displays the data"| viewer
agent <-->|"questions, answers"| you
1. Install on macOS or Linux:
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/MBagory/wise-bucket/releases/latest/download/wisebucket-installer.sh | sh
Then open a new terminal. To update, run it again. To build from source instead, see the guide.
2. Set up (once): wisebucket setup asks which folders hold your recordings.
3. Connect your agent (once) with the command setup prints. Claude Code:
claude mcp add --scope user wise-bucket -- ~/.cargo/bin/wisebucket serve
Other agents (Codex, Cursor, Cline…): see Connect your agent.
4. Ask, in a new agent session:
you › In the demo ULog, what is the accelerometer noise floor at rest?
agent › ⏺ wise-bucket · sql
SELECT count(*), stddev("accelerometer_m_s2[2]"), avg("accelerometer_m_s2[2]")
FROM stream('sensor_combined')
Over 17,070 samples, the vertical acceleration averages −9.62 m/s²
(gravity) with a standard deviation of 0.16 m/s²: that is the noise
floor at rest.
you › Show me the three accelerometer axes in Rerun.
agent › ⏺ wise-bucket · rerun (entity: sensor_combined)
SELECT t, "accelerometer_m_s2[0]" AS x, "accelerometer_m_s2[1]" AS y,
"accelerometer_m_s2[2]" AS z
FROM stream('sensor_combined')
Sent 17,070 rows to the Rerun viewer: sensor_combined/x, y and z,
on the timeline t.
The Rerun viewer must be running (rerun). Wise Bucket sends the rows to it directly; they do not pass through your agent.
wisebucket demo --all # about 4 MB of public samples, declared as the `demo` folder
wisebucket demo --list # formats, sizes and licenses
Samples are fetched from their upstream projects at pinned commits, SHA-256-verified, and stay under their own licenses. Then ask your agent:
Runs on macOS and Linux; Windows through WSL2.
User guide · Troubleshooting and FAQ · Reference: tables, CLI, configuration, errors · Security and privacy · Contributing
Apache License 2.0. Built on Apache DataFusion, Apache Parquet, Apache Arrow, the Rust MCP SDK, mcap, Rerun's re_ros_msg, px4-ulog, rust-mavlink and SQLite. Bundled ROS 2 message definitions come from ros2/common_interfaces and related repositories (Apache-2.0, BSD-3-Clause). To cite Wise Bucket, use CITATION.cff.
Long-term context and insights on your time-series data: an MCP server that lets your AI agent query all your boat, robot, drone and vehicle logs with SQL, on your machine. NMEA 0183/2000, ROS 2, MCAP
總覽
Long-term context and insights on your time-series data: an MCP server that lets your AI agent query all your boat, robot, drone and vehicle logs with SQL, on your machine. NMEA 0183/2000, ROS 2, MCAP, PX4, ArduPilot, MAVLink, CAN with DBC, MDF4.
工具
尚無工具清單。遠端伺服器會在首次健康檢查後取得工具清單。
常見問題
在本頁複製適用於 Cursor、Claude Desktop、Claude Code 或 VS Code 的 stdio 設定。它使用發布者的安裝方式,而不是 PlugMCP 代理伺服器。
這筆記錄未列出必要的環境變數。
Wise Bucket 透過 stdio 在本機執行。
上次檢查時間:2026-10-28。 Long-term context and insights on your time-series data: an MCP server that lets your AI agent query all your boat, robot, drone and vehicle logs with SQL, on your machine. NMEA 0183/2000, ROS 2, MCAP, PX4, ArduPilot, MAVLink, CAN with DBC, MDF4.