canforge-mcp
OfficialAllows ChatGPT to inspect DBC files and decode CAN capture logs locally via a Secure MCP Tunnel.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@canforge-mcp decode frame 0x100#00AA from capture.asc"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
canforge-mcp
canforge-mcp is a local, read-only MCP server for inspecting DBC files and
decoding CAN capture logs.
Files stay on the machine running the server. The server exposes bounded tools instead of uploading captures or returning unbounded traces.
Tools
Tool | Purpose |
| DBC version, message/signal/node counts, and node names |
| Bounded message summaries, with optional search |
| Full message and signal detail by name or arbitration ID |
| Bounded signal search across a DBC |
| Decode one hexadecimal CAN payload |
| Structured DBC validation issues |
| Added, removed, and changed messages and signals |
| Detect a log format and read header metadata |
| Frame count, span, ID counts, and median cycle times |
| One-pass inventory of DBC signals observed in a log |
| Bounded raw-frame samples with ID and time filters |
| Bounded decoded frames from a DBC and log |
| Downsampled timestamp/value points for one signal |
See the tool reference for arguments, return shapes, and hard caps.
Related MCP server: telemetry-mcp
Install
Run directly with uvx:
uvx canforge-mcpOr install with pip:
pip install canforge-mcp
canforge-mcpRequires Python >=3.11.
Configure
Claude Code:
claude mcp add canforge -- uvx canforge-mcpClaude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"canforge": {
"command": "uvx",
"args": ["canforge-mcp"]
}
}
}Restart Claude after changing its MCP configuration.
ChatGPT (Secure MCP Tunnel)
ChatGPT cannot start a local stdio MCP server directly. Use OpenAI's Secure MCP Tunnel to keep Canforge running locally without exposing it to the public internet.
Before starting, enable developer mode in ChatGPT under Settings → Security
and login, then create a tunnel in the OpenAI Platform. You need its tunnel
ID, a runtime API key, and the tunnel-client binary. Make sure the tunnel is
associated with the ChatGPT workspace where you will use Canforge.
Configure and start the tunnel with placeholder credentials:
export CONTROL_PLANE_API_KEY="sk-..."
tunnel-client init \
--sample sample_mcp_stdio_local \
--profile canforge \
--tunnel-id tunnel_your_id \
--mcp-command "uvx canforge-mcp"
tunnel-client doctor --profile canforge --explain
tunnel-client run --profile canforgeKeep tunnel-client run running while using Canforge. In ChatGPT, open
Settings → Plugins, add a developer-mode app, choose Tunnel as the
connection, and select or paste the tunnel ID. Add the new app to a chat before
asking ChatGPT to use the Canforge tools.
Canforge resolves paths on the machine running tunnel-client. Files attached
directly to a ChatGPT conversation are not automatically available as local
filesystem paths; provide an accessible local path instead.
Design
Local-first: tools accept filesystem paths and do not send file content over the network.
Read-only: no tool creates, edits, encodes, or overwrites a file.
Bounded: list and frame tools enforce hard caps and report
total,returned, andtruncated; timeseries are downsampled server-side.Cached: parsed DBCs are cached by resolved path and nanosecond modification time for repeated inspection during one server session.
Composable: capkit reads capture formats; dbckit parses and decodes DBC content.
Stdio-only: the 0.x line exposes no network transport or hosted service.
Scope and Caveats
Supported capture formats come from capkit 0.3: Kvaser CanKing TXT, candump text, and Vector ASC.
DBC support and validation behavior follow dbckit 1.x.
Timestamps are floats exactly as recorded by capkit; they are not rebased.
Median cycle time is the median gap between consecutive occurrences of an ID.
capkit adds raw priority, PGN, and source-address fields for observed extended IDs; dbckit remains responsible for DBC-aware J1939 matching and decoding.
signal_timeseriesuses deterministic, evenly spaced index sampling when a series exceedsmax_points; it is intended for inspection, not resampling or signal processing.log_signal_inventoryalways loads the DBC leniently and reports parse diagnostics and per-message decode safety. It scans the log body once;include_values=trueadds bounded distinct decoded values to the inventory.Paths are resolved by the machine running the MCP server. A remote client's filesystem is not visible to a server running elsewhere.
Development
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
ruff check .
mypy canforge_mcp
pytest --cov=canforge_mcp --cov-fail-under=90
python -m buildLarge-log performance checks are opt-in and generate their own deterministic captures. See benchmarks/README.md for the smoke command, the full 100k/1M matrix, measurement policy, and committed v0.2.0 baseline.
License
MIT
Available Tools
13 toolsdbc_infoARead-onlyIdempotent
Summarize a local DBC file. Use this first to learn its size and node names.
Inputs: dbc_path (required). Returns: object
decode_frameARead-onlyIdempotent
Decode one CAN payload using a local DBC. data_hex is a hexadecimal byte string; no file is modified.
Inputs: data_hex (required), dbc_path (required), arbitration_id (required). Returns: object
decode_logARead-onlyIdempotent
Decode a bounded sample of local log frames through a local DBC. Filters combine by intersection; limit is capped at 500.
Inputs: dbc_path (required), log_path (required), limit, messages, +3 more. Returns: object
diff_dbcsARead-onlyIdempotent
Compare two local DBC files and return bounded added, removed, and changed message/signal details.
Inputs: dbc_a_path (required), dbc_b_path (required). Returns: object
get_messageARead-onlyIdempotent
Return one message and its signal layout. message accepts a name, decimal ID, or 0x-prefixed hex ID.
Inputs: message (required), dbc_path (required). Returns: object
list_messagesARead-onlyIdempotent
List DBC messages, optionally narrowing by name, ID, or sender. The server caps limit at 200.
Inputs: dbc_path (required), limit, search. Returns: object
log_signal_inventoryBRead-onlyIdempotent
Inventory the DBC signals observed in one full, one-pass scan of a local CAN log.
Inputs: dbc_path (required), log_path (required), match_mode, include_values. Returns: object
log_statsARead-onlyIdempotent
Scan a local CAN log for counts, span, top IDs, and median per-ID cycle times. top is capped at 200.
Inputs: log_path (required), top. Returns: object
probe_logARead-onlyIdempotent
Detect a local CAN log's format and cheap header metadata without scanning all frames.
Inputs: log_path (required). Returns: object
read_framesARead-onlyIdempotent
Read a bounded raw-frame sample from a local log, with optional ID and timestamp filters. limit is capped at 500.
Inputs: log_path (required), limit, time_end, id_filter, +1 more. Returns: object
search_signalsARead-onlyIdempotent
Search signal names and comments across a DBC. The server caps limit at 200.
Inputs: query (required), dbc_path (required), limit. Returns: object
signal_timeseriesARead-onlyIdempotent
Return a bounded, evenly downsampled [timestamp, value] series for one DBC signal. max_points is capped at 2,000.
Inputs: signal (required), message (required), dbc_path (required), log_path (required), +3 more. Returns: object
validate_dbcARead-onlyIdempotent
Validate a DBC and return structured issues. The server caps limit at 200.
Inputs: dbc_path (required), limit. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 13 tools
Each tool targets a distinct operation on DBC or log files. Even the multiple log-analysis tools (probe_log, log_stats, read_frames, decode_log) have clearly defined differences in scope and output.
Most tools follow verb_noun pattern (validate_dbc, list_messages, decode_log), but a few are noun-led (dbc_info, log_stats) or compound nouns (signal_timeseries), creating minor inconsistency.
13 tools is well within the ideal range, each covering a unique aspect of DBC and CAN log processing without redundancy.
The set comprehensively covers DBC validation, informational retrieval, message/signal access, comparison, and log probing, scanning, decoding, and timeseries extraction. No obvious missing operations for this domain.
Maintenance
Related MCP Connectors
An MCP server for search and retrieval over a local LanceDB database
Read-only MCP server for The Quiet Protocol's engines, benchmarks, proof, and business data.
Self-hosted MCP server: 26 deterministic dev, security, and EVM tools.
Read-only MCP server for the WebAssembly spec: instructions, types, sections, search, proposals.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceLocal-first MCP server for querying multi-repo engineering documentation artifacts from a SQLite corpus.3 npmAGPL 3.0
- AlicenseNot gradedqualityCmaintenanceA read-only MCP server for querying telemetry data from configurable backends. Provides tools to list sources, describe schemas, run bounded queries, and compute aggregates.MIT
- AlicenseAqualityDmaintenanceA Model Context Protocol (MCP) server that exposes lnav log file analysis capabilities to AI assistants, specifically optimized for Kvaser Plain Text Log Frame CAN bus log processing.9MIT
- AlicenseAqualityCmaintenanceA local-first, read-only MCP server for vehicle diagnostics that exposes structured OBD-II tools for reading PIDs, DTCs, and ECU snapshots without raw protocol access.71Apache 2.0