Ever wondered why your robot died with 30 seconds left? Or why auto worked in practice but not in competition?
wpilog-mcp lets you ask those questions in plain English. Load your robot's telemetry logs and have a conversation with your data.
The Model Context Protocol (MCP) is an open standard that enables AI assistants (like Claude) to securely access your local data and tools.
This project provides an MCP Server. Once configured, it acts as a "bridge" that gives your AI assistant the specific tools needed to read WPILOG files, analyze swerve performance, detect brownouts, and even pull match results from The Blue Alliance—all through a natural conversation.
Unlike traditional log viewers (like AdvantageScope) which require you to know exactly what to look for, wpilog-mcp allows you to ask high-level engineering and strategic questions. The AI doesn't just "query" data; it hypothesizes, investigates, and synthesizes.
Prompt: "We just finished Q68 and the drivers said the robot 'stuttered' during teleop. Investigate the log and tell the pit crew exactly what to check."
- The AI's Reasoning: Claude will load the log, scan the
get_ds_timelinefor brownout events, usepower_analysisto find which motor controller had the highest current spike at that exact timestamp, and checkcan_healthfor timeouts. - The Result: "I found a BROWNOUT_START at 42.5s. During this time, the 'Intake/Roller' current spiked to 60A while velocity was zero, suggesting a mechanical jam. Check the intake for debris or a bent mounting bracket."
Prompt: "Compare our cycle times in Q74 vs Q68. Why were we slower in the second half of Q68?"
- The AI's Reasoning: Claude will pull match results from TBA to see the scores, use
analyze_cyclesto calculate state-based efficiency, and correlate "dead time" with robot position data. - The Result: "Your scoring cycles in Q74 averaged 8.2s. In Q68, they slowed to 12.5s after the 60-second mark. I noticed that during those slower cycles, the robot was taking a much longer path around the 'Stage' obstacle—check if your autonomous path-finding or driver path was blocked."
Prompt: "Look at our swerve drive performance in the last match. Is our steering PID too aggressive? Look for oscillation."
- The AI's Reasoning: Claude will use
analyze_swerveto identify the modules, callget_statisticson the steering error, and runfind_peaksto look for high-frequency oscillations in theAppliedVolts. - The Result: "The Back-Left module is showing a 0.15s oscillation period in steering position while the robot is at a standstill. This suggests your P gain is slightly too high or your D gain is insufficient for the new modules."
Built by FRC Team 2363 Triple Helix using WPILib's official DataLogReader for guaranteed format compatibility.
See what's possible: Example Analysis Report - A complete match analysis generated from real robot logs.
Built for JDK 17 binary compatibility. It is recommended to use the WPILib JDK, which the build tool attempts to locate automatically.
git clone https://github.com/TripleHelixProgramming/wpilog-mcp.git
cd wpilog-mcp
./gradlew shadowJarConfiguration location depends on how you're running Claude Code:
| Environment | Config File |
|---|---|
| VS Code extension | .mcp.json in your project folder |
| Claude Code CLI | ~/.claude/settings.json |
| Claude Desktop | See Claude Desktop section |
Example configuration (macOS/Linux):
{
"mcpServers": {
"wpilog": {
"command": "/path/to/wpilog-mcp/run-mcp.sh",
"args": [
"-logdir", "/path/to/your/logs",
"-team", "2363"
],
"env": {
"TBA_API_KEY": "your_tba_api_key"
}
}
}
}Example configuration (Windows):
{
"mcpServers": {
"wpilog": {
"command": "C:\\path\\to\\wpilog-mcp\\run-mcp.bat",
"args": [
"-logdir", "C:\\path\\to\\your\\logs",
"-team", "2363"
],
"env": {
"TBA_API_KEY": "your_tba_api_key"
}
}
}
}The wrapper scripts (run-mcp.sh / run-mcp.bat) automatically locate the WPILib JDK and configure JVM memory. To control how much memory the server can use, set the WPILOG_MAX_HEAP environment variable (default: 4g):
{
"mcpServers": {
"wpilog": {
"command": "/path/to/wpilog-mcp/run-mcp.sh",
"args": ["-logdir", "/path/to/your/logs"],
"env": {
"WPILOG_MAX_HEAP": "8g",
"TBA_API_KEY": "your_tba_api_key"
}
}
}
}Note: TBA key is optional. Without it, the server works normally but without match enrichment. Team number is extracted from log metadata when available, with -team as a fallback.
Restart Claude Code. Then just ask:
What robot logs are available?
Load the qualification match 42 log and give me a summary
When did battery voltage drop below 11 volts?
Compare the commanded wheel speeds to actual wheel speeds
That's it! See Usage Examples for more.
- Quick Start
- Usage Examples
- Configuration Options
- Available Tools
- Supported Data Types
- Troubleshooting
- Development
- Contributing
What robot logs are available?
The server parses filenames like 2024vadc_qm42.wpilog into friendly names like "VADC Qualification 42".
Load the robot log and give me a summary
Show me all entries related to the drivetrain
What Pose2d entries are available?
Show me where the robot was at the start and end of the match
Get statistics on the robot's X position - how much did it vary?
When did battery voltage drop below 11 volts?
Find all times when robot speed exceeded 4 meters per second
Search for any CAN errors in the console output
Find all vision-related messages in the logs
Show me the swerve module states for the front left module
Compare commanded wheel speeds to actual wheel speeds
Compare /RealOutputs/Drive/Pose with /ReplayOutputs/Drive/Pose
Load both the practice and match logs, then compare battery statistics
What REV signals are available for this match?
Show me the motor temperature and current for SparkMax_1 during teleop
Compare the commanded output from the wpilog with the actual applied output from the revlog
Unload all logs to free up memory
| Option | Description |
|---|---|
-logdir <path> |
Directory containing .wpilog files (scans subdirectories up to 3 levels deep) |
-team <number> |
Default team number for logs missing metadata |
-tba-key <key> |
The Blue Alliance API key for match data enrichment |
-maxlogs <n> |
Max number of logs to cache (default: 20) |
-maxmemory <mb> |
Max memory (MB) for log cache (alternative to -maxlogs) |
-debug |
Enable debug logging |
-help |
Show usage information |
Environment variables:
WPILOG_DIR- Alternative to-logdir(command line takes precedence)WPILOG_TEAM- Alternative to-teamTBA_API_KEY- Alternative to-tba-keyWPILOG_MAX_HEAP- JVM max heap size (default:4g). Set in the MCPenvblock or export before running the wrapper script. Examples:2g,8g,512m
Cache limits: The server caches parsed logs in memory for fast access. By default, up to 20 logs are kept. Use -maxlogs to change this limit, or -maxmemory to set a memory-based limit instead (only used if -maxlogs is not set). The least recently used log is evicted when the limit is reached.
JVM memory: The wrapper scripts (run-mcp.sh / run-mcp.bat) set the JVM heap to 4 GB by default. Large log files (hundreds of MB) may need more heap. Set WPILOG_MAX_HEAP=8g for heavy use. The server will refuse to load files that would exceed available heap rather than crashing.
Get a free API key at thebluealliance.com/account.
When configured, the server enriches match logs with TBA data:
- Match times - Corrects midnight timestamps from FMS
- Scores - Your alliance's score and opponent's score
- Win/Loss - Whether your team won the match
- Alliance - Which alliance (red/blue) your team was on
Team number is extracted from each log file's metadata (DriverStation/FMS data), with the configured -team value as a fallback.
This data is automatically added to list_available_logs output for logs that have event/match/team metadata.
wpilog-mcp can correlate .revlog files (from REV's logging library on the roboRIO, or REV Hardware Client on a laptop) with your robot's WPILOG data, giving you access to high-resolution motor controller telemetry with synchronized timestamps.
How it works:
- Place
.revlogfiles in the same directory as their corresponding.wpilogfiles - Load the wpilog normally — revlog files are discovered and synchronized automatically
- Use
sync_statusto verify synchronization confidence before relying on timestamps
Synchronization: Timestamps are aligned using a two-phase approach:
- Coarse alignment from
systemTimeentries and revlog filename timestamps (seconds-level) - Fine alignment via Pearson cross-correlation of matching signals like motor output duty cycle (millisecond-level)
For long recordings (>15 minutes), linear clock drift between the FPGA clock and the monotonic clock is estimated and compensated automatically.
The system reports confidence levels (HIGH/MEDIUM/LOW/FAILED) based on correlation strength, number of agreeing signal pairs, and inter-pair consistency. Typical accuracy at HIGH confidence is ±1–5 ms.
If automatic sync fails, use set_revlog_offset to manually provide a known offset.
Limitations:
- REV logs use
CLOCK_MONOTONICwhile WPILOGs use FPGA time — offset varies per boot - Correlation requires overlapping signal variation (flat/disabled data degrades quality)
- Short logs or steady-state data may produce lower confidence synchronization
Available data:
- Applied output (duty cycle), velocity, position
- Bus voltage, output current, temperature
- Faults and sticky faults
See TOOLS.md for detailed tool documentation and a technical explanation of the synchronization algorithm.
Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):
{
"mcpServers": {
"wpilog": {
"command": "/path/to/wpilog-mcp/run-mcp.sh",
"args": ["-logdir", "/path/to/logs"]
}
}
}This server uses stdio transport (JSON-RPC over stdin/stdout). Generic config:
{
"command": "/path/to/wpilog-mcp/run-mcp.sh",
"args": [],
"transport": "stdio"
}See MCP Protocol for client implementations.
wpilog-mcp provides 46 tools organized into categories:
| Category | Tools |
|---|---|
| Core | load_log, list_entries, read_entry, get_entry_info, search_entries, get_statistics, compare_entries, get_types, list_struct_types, health_check |
| Log Browser | list_available_logs |
| Multi-Log | list_loaded_logs, set_active_log, unload_log, unload_all_logs |
| Search | find_condition, search_strings |
| Analysis | detect_anomalies, find_peaks, rate_of_change, time_correlate, get_match_phases |
| FRC-Specific | analyze_swerve, power_analysis, can_health, compare_matches, get_code_metadata |
| FRC Domain Analysis | get_ds_timeline, analyze_vision, profile_mechanism, analyze_auto, analyze_cycles, analyze_replay_drift, analyze_loop_timing, analyze_can_bus |
| TBA Integration | get_tba_status |
| RevLog Integration | list_revlog_signals, get_revlog_data, sync_status, set_revlog_offset |
| Export | export_csv, generate_report |
- Execution Time Tracking: All tool responses include
_execution_time_msfield for performance monitoring - Intelligent Error Handling: Clear error messages with specific error codes and "Did You Mean?" suggestions for misspelled tool names
- Memory Monitoring: Real-time heap usage tracking with estimation accuracy validation via
health_checktool - Improved Statistical Analysis: IQR calculation uses proper linear percentile interpolation for accurate outlier detection
- Enhanced Mechanism Analysis: Stall detection, settling time, and overshoot calculations for control system tuning
- Vision System Monitoring: Pose jump detection to identify unreliable vision estimates
- Loop Timing Analysis: Detect and diagnose real-time performance issues
- CAN Bus Health: Monitor bus utilization and error rates
- REV Log Integration: Correlate high-resolution motor controller data from
.revlogfiles with FPGA-timestamped telemetry
For complete tool documentation with parameters and examples, see TOOLS.md.
boolean, int64, float, double, string, raw, json, and arrays of each
| Type | Decoded Fields |
|---|---|
Pose2d |
x, y, rotation_rad, rotation_deg |
Pose3d |
x, y, z, qw, qx, qy, qz |
Translation2d/3d |
x, y, (z) |
Rotation2d |
radians, degrees |
Rotation3d |
qw, qx, qy, qz |
Transform2d/3d |
Same as Pose |
Twist2d/3d |
dx, dy, (dz), dtheta/(rx, ry, rz) |
| Type | Decoded Fields |
|---|---|
ChassisSpeeds |
vx_mps, vy_mps, omega_radps |
SwerveModuleState |
speed_mps, angle_rad, angle_deg |
SwerveModulePosition |
distance_m, angle_rad, angle_deg |
| Type | Decoded Fields |
|---|---|
TargetObservation |
yaw_rad, yaw_deg, pitch_rad, pitch_deg, skew_rad, skew_deg, area, confidence, objectID |
PoseObservation |
timestamp, pose_x, pose_y, pose_z, pose_qw, pose_qx, pose_qy, pose_qz, ambiguity, tagCount, averageTagDistance, type |
SwerveSample |
timestamp, x, y, heading, heading_deg, vx, vy, omega, ax, ay, alpha, moduleForcesX[4], moduleForcesY[4] |
Note: PoseObservation.type is decoded as an enum string: MEGATAG_1, MEGATAG_2, or PHOTONVISION.
-
Check Java version - Requires JDK 17 binary compatibility (JDK 21 is also fine):
/path/to/wpilib/2026/jdk/bin/java -version # Should show: openjdk version "17.x.x" or "21.x.x" -
Test manually:
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{}}' | \ ./run-mcp.sh
-
Check config location:
- VS Code extension:
.mcp.jsonin your project folder - Claude Code CLI:
~/.claude/settings.json - Claude Desktop:
~/Library/Application Support/Claude/claude_desktop_config.json
- VS Code extension:
-
Restart completely - Quit and relaunch (not just reload)
Truncated logs (from robot power loss) are handled gracefully - the server recovers as much data as possible and marks the log as truncated.
Usually means wrong Java version or incorrect path. Verify the JAR exists at the configured path.
- JDK 17+ (WPILib 2026 JDK recommended)
- No other dependencies (WPILib libraries bundled)
WPILib JDK locations:
- macOS:
~/wpilib/2026/jdk/bin/java - Windows:
C:\Users\Public\wpilib\2026\jdk\bin\java.exe - Linux:
~/wpilib/2026/jdk/bin/java
wpilog-mcp/
├── src/main/java/org/triplehelix/wpilogmcp/
│ ├── Main.java # Entry point
│ ├── log/
│ │ ├── LogManager.java # WPILOG parsing & struct decoding
│ │ └── LogDirectory.java # Log file discovery
│ ├── mcp/
│ │ ├── McpServer.java # MCP protocol handling
│ │ └── JsonRpc.java # JSON-RPC utilities
│ ├── revlog/
│ │ ├── RevLogReader.java # REV log file parsing
│ │ └── RevLogSignal.java # Signal data structures
│ ├── sync/
│ │ ├── LogSynchronizer.java # Cross-correlation timestamp alignment
│ │ ├── SynchronizedLogs.java # Unified wpilog+revlog access
│ │ └── SyncResult.java # Synchronization confidence metrics
│ ├── tba/
│ │ ├── TbaClient.java # TBA API client with caching
│ │ ├── TbaConfig.java # TBA configuration management
│ │ └── TbaEnrichment.java # Log enrichment with TBA data
│ └── tools/
│ ├── WpilogTools.java # Tool registration
│ ├── CoreTools.java # Log management tools (11)
│ ├── QueryTools.java # Search & query tools (4)
│ ├── StatisticsTools.java # Statistical analysis tools (6)
│ ├── RobotAnalysisTools.java # FRC analysis tools (7)
│ ├── FrcDomainTools.java # Advanced FRC tools (9)
│ ├── ExportTools.java # CSV/report export (2)
│ ├── TbaTools.java # TBA integration (1)
│ ├── RevLogTools.java # REV log integration (4)
│ └── ToolUtils.java # Shared utilities
├── src/test/java/ # Test suite
├── build.gradle # Build configuration
├── README.md
└── TOOLS.md # Complete tool reference
./gradlew build # Full build with tests
./gradlew shadowJar # Build fat JAR only
./gradlew test # Run tests- Report bugs - Open an issue with reproduction steps
- Request features - Open an issue describing your use case
- Submit PRs - Fork, make changes, add tests, submit
MIT License - see LICENSE
- WPILib for the DataLog format
- AdvantageKit for pioneering FRC replay logging
- Anthropic for MCP and Claude
- FRC Team 2363 Triple Helix
- TOOLS.md - Complete tool reference
- EXAMPLE.md - Sample analysis report showing tool capabilities
- WPILib DataLog Docs
- MCP Protocol