embedcalc-mcp

Exact embedded-engineering math for AI agents. An MCP server with 29 deterministic calculators and code generators — so Claude, Cursor, or any MCP-enabled agent computes CRC check values, CAN timing registers, and struct padding instead of hallucinating them.

Why

LLMs are confidently wrong about exactly the math that bricks boards. Ask any model for a CRC-16/MODBUS check value, an STM32 CAN_BTR register for 500 kbit/s @ 36 MHz, or the padded size of a struct — you'll get a plausible-looking wrong answer often enough that you can't trust any of them. The failure mode is the worst kind: the answer looks right.

This server gives the model tools instead. Every tool is a pure, deterministic implementation pinned with test vectors (standard CRC check values, oracle fixtures, IPC/ASTM published numbers — 99 tests).

Install

Claude Code:

claude mcp add embedcalc -- npx -y embedcalc-mcp

Claude Desktop / Cursor / any MCP client (JSON config):

{
  "mcpServers": {
    "embedcalc": { "command": "npx", "args": ["-y", "embedcalc-mcp"] }
  }
}

No API key, no account. Runs locally over stdio. (The optional find_component tool queries embedcalc.com over the network; everything else is offline.)

Tools (29)

Domain	Tools
Code generation	crc_to_c (7 presets + custom Rocksoft params → compilable table-driven C with check value) · regmap_to_c (LSB-first fields → shift/mask defines + bit-field struct) · bitfield_macro
Timing	can_bit_timing (bxCAN BTR / FDCAN NBTP / MCP2515 CNF1-3) · uart_baud (STM32 BRR + error %) · i2c_timing (pull-up range + STM32 TIMINGR) · spi_timing · i2s_clocks · lin_timing · rs485_bus · modbus_rtu_timing (t1.5/t3.5) · ring_buffer_size
Data representation	qformat_convert (Qm.n fixed point) · ieee754_convert (float/double, hex↔decimal↔bin) · number_base_convert (32-bit two's complement) · endian_swap (incl. Modbus BADC/CDAB) · struct_layout (C parser → offsets/padding/size, 32-bit MCU model)
Electronics	eseries_nearest (E6/E12/E24/E96) · resistor_code (color bands + SMD/EIA-96) · crystal_load_caps · lm317 · attenuator_pad (T/π) · junction_temp · microstrip_impedance (IPC-2141) · pcb_trace_width (IPC-2221) · awg_wire · ne555 · dcdc_converter (buck/boost/inverting)
Components	find_component — real, in-stock MPNs + buy links from a computed value (parametric search)

All tool names are prefixed embedcalc_ (e.g. embedcalc_crc_to_c).

Example prompts

"Give me MODBUS CRC-16 as a C function" → table-driven C, check("123456789") = 0x4B37 embedded for on-target verification

"CAN 500 kbit/s on STM32F103, APB1 36 MHz, sample point 87.5%" → CAN_BTR = 0x00050008 (BRP=9, TS1=6, TS2=1)

"How big is struct { uint8_t flags; uint32_t ts; uint16_t id; } on a Cortex-M?" → 12 bytes, 5 of them padding, with the layout table

"Encode −0.25 in Q1.14" / "What float is 0x7FC00000?" / "Nearest E96 to 12.3 kΩ?"

Accuracy

• CRC: all 7 presets verified against standard check values for "123456789"
• CAN: solver output byte-identical to the interactive calculators at embedcalc.com (oracle fixtures in tests/fixtures/)
• IEEE-754 / AWG / IPC: pinned to reference constants and published standard values
• struct_layout uses a 32-bit MCU model (pointers 4 B; double/int64 size 8, align 4) — stated in the tool description so the model can caveat it

Run it yourself: npm install && npm run build && npm test (99 tests, includes a stdio end-to-end pass that spawns the built server and calls the tools).

Related

• embedcalc.com — the same 60+ calculators as free interactive tools, no AI required
• Found a calculation LLMs keep getting wrong? Open an issue — collecting candidates for the next batch.

License

MIT