openm454

English | 简体中文

An open driver, command-line interface and MCP server for the DreamSourceLab DSCope M454 USB mixed-signal oscilloscope (4 analog channels up to 400 MSa/s, 8 digital channels, built-in 0-3 V signal generator), written so that an AI agent (Claude Code, or any MCP client) can capture and analyze real signals by itself, without the vendor GUI.

The M454 speaks a newer ("v3") USB protocol than the devices supported by the open-source DSView. The protocol was recovered from DSWave 1.2.1 by static analysis, by driving DSWave's own hardware layer through injected commands and by USB tracing; see docs/PROTOCOL.md.

Use

Requirements: Windows, Python 3.10+, the scope bound to WinUSB (it is by default), a USB 3 port and cable (captures refuse a USB 2 link with UNSUPPORTED), and DSWave 1.2.1, needed once to extract the FPGA bitstream (it is never redistributed). Close DSWave before using the driver: the scope accepts one USB client.

pip install openm454            # or, from a clone: pip install -e .

The openm454 data folder is the clone itself, or for an installed copy %LOCALAPPDATA%\openm454 (~/.local/share/openm454 elsewhere; OPENM454_HOME overrides it). It holds the cached bitstream (firmware/), state (state/) and captures (captures/).

Where the FPGA bitstream comes from

DSWave 1.2.1 (DSWave.exe) is needed once. openm454 only reads its files, looks in these places in order (each may be the folder holding DSWave.exe or the exe itself) and searches no disk beyond them:

1. the environment variable OPENM454_DSWAVE (DSCOPE_DSWAVE is accepted too);
2. the path saved in state/dswave.json of the data folder (written by openm454 setup --dswave, below);
3. DSWave/ in the data folder (a copy of the DSWave folder);
4. the Windows uninstall entries of DSWave in the registry: a normal DSWave installation is found with no setup at all;
5. %ProgramFiles%, %ProgramFiles(x86)% and %LOCALAPPDATA%\Programs, each in DSWave\ or DreamSourceLab\DSWave\.

The bitstream is extracted the first time the scope's FPGA has to be configured (the first command after the scope is powered up) or by setup --dswave, checked by SHA-256 and cached in firmware/ of the data folder; DSWave can then be closed, moved or uninstalled. setup without --dswave only reports (cached, dswave, source, ready). If nothing is found, the error is FIRMWARE_SOURCE (exit 6) with what was searched and what to do. To see the state, or to tell openm454 where DSWave is once (no scope needed):

python -m openm454 setup                          # cached? where is DSWave found, and how
python -m openm454 setup --dswave <folder>        # check DSWave 1.2.1 there, cache the bitstream, save the path

The saved path is read, with no restart, by every openm454 process that uses the same data folder (the MCP server included: if it was registered with -e OPENM454_HOME=<folder>, run setup with the same OPENM454_HOME).

python -m openm454 info
python -m openm454 capture -c A1,D0 --vdiv 1V --offset 1.5V --timebase 200us --trig A1:rise:1.4V
python -m openm454 watch -c A1 --every 0.5s --count 10 --fields A1.freq_hz,A1.vpp_v
python -m openm454 awg set --wave sine --freq 10k --low 0V --high 2V
python -m openm454 analyze captures/<file>.npz --window=-1ms:1ms
python -m openm454 capture -c D0,D1 --rate 20M --duration 2ms --trig D0:pulse>:200us --decode "i2c:scl=D0,sda=D1"
python -m openm454 selftest

Recording modes and measured speeds

Measured on the M454 over USB 3 with every input on a 1 kHz square, every period checked for gaps (2026-11-01/02):

• The data rate is rate x (analog channels + digital channels / 8) bytes/s.
• stream: real time at DSWave's 100 MB/s ceiling and at 200 MB/s. At 400 MB/s the readout ran at 270-290 MB/s and the rest waited in the sample memory; the record arrived complete. 12 channels at 50 MSa/s (250 MB/s, ~2 s records): 0.74 / 0.66 s between records when sent while acquired (the default), 2.35 / 2.45 s with --record-first; records complete and continuous.
• log: 150 MB/s (2 analog + 8 digital at 50 MSa/s) for 3 min (27 GB) and for 30 s to disk with a constant host lag; 200 MB/s (2 analog at 100 MSa/s) for 20 s without backlog; 5 MB/s for 10 min. At 250 MB/s (4 analog + 8 digital at 50 MSa/s) the readout reached ~200 MB/s and the backlog grew ~55 MB/s. A 300 MB/s burst of 3 s (2 analog + 8 digital at 100 MSa/s) read out at 266 MB/s and stayed complete.
• log above 100 MSa/s interleaves the converters exactly as capture does (DSWave never does this in loop mode). Verified on the scope: 1 analog channel at 200 and 400 MSa/s, 2 analog at 200, 2 digital at 400, 4 digital at 200 and mixed sets. 2 digital channels at 400 MSa/s or 4 at 200 MSa/s are 100 MB/s and run without a time limit. 1 analog channel at 200 MSa/s (200 MB/s) ran 10 s without backlog.
• The sample memory is a 512 MiB ring in loop mode: the host may fall up to ~512 MiB behind without losing data (host stalls of 10 s at 50 MB/s were absorbed). Beyond that the scope overwrites unread data without any flag and keeps streaming: from then on the data are garbage (one phase step on analog, the digital channels misaligned) until the run is restarted. Measured with the driver's backlog estimate, the break came at 508-514 MiB (three runs), so log stops at 480 MiB.
• log --plan (no device needed) prints the verdict for a rate, the gap-free limit and recommendation (next section), how long the run can last (max_duration_s) and where the data goes: RAM by default (refused when it does not fit the budget unless --keep-last), --storage disk for ~256 MB chunk files (it warns about SSD wear: 200 MB/s is 720 GB per hour), --storage none for statistics only.
• bench measures this PC (free RAM, decode speed, disk write speed, USB readout) and records how much RAM and disk the owner allows for recordings and the data folder (bench --ram 8GB --disk 50GB --dir <folder>, state/host.json); log takes its defaults from there. Measured here: decode 468 MB/s, USB readout 206 MB/s (4 analog + 8 digital) / 267 MB/s (2 analog + 8 digital).

Above the readout limit: how log decides

D = the data rate; R = how fast this PC reads the scope while it acquires. R depends on the channel set (measured here: 202 MB/s for 4 analog + 8 digital at 50 MSa/s, 264 MB/s for the same at 100 MSa/s and for 2 analog + 8 digital at 100 MSa/s, 275 MB/s for 4 analog at 100 MSa/s).

1. R is measured, not assumed: the first log above 150 MB/s with a channel set runs a probe of up to 2 s (a probe line) and remembers R in state/host.json: the measured rate, or "kept up" (R >= D) when it reads at

   = 99 % of D. Later runs above 150 MB/s refresh it when the result is clear (still falling behind at a steady rate at the stop, or caught up at the end), and so do bench and --remeasure. Until then --plan assumes 190 MB/s and marks the plan provisional.

2. D <= R: the host keeps up (sustained / near_limit); the run lasts as long as the storage allows.
3. D > R: the scope gets D - R ahead per second (backlog_growth_mb_s), and it sends nothing after a stop, so the unread part is lost. A log is gap-free for at most gapfree_max_s = 480 MiB x R / (D x (D - R)), the signal the host has read when the scope is 480 MiB ahead. A shorter log runs normally.
4. A longer log (or forever) is compared with what stream would do with the same channels and settings (stream_alternative, from stream's own planning): records of up to 400M samples per channel / 512 MiB, sent while acquired, with gaps of record x (D - R) / R + 0.18 s, at stream's rates (1-2-5 steps x interleave; it may run a domain slower, e.g. 2 analog + 8 digital at 100 MSa/s samples the digital channels at 50 MSa/s):
   • STREAM_BETTER (exit 2): stream samples every channel at least as fast, and its records are at least as long as the log's one gap-free piece (or one record holds the whole duration). The error carries the stream command, with the same vertical settings.
   • NEEDS_FORCE (exit 2): the log's piece is the longest gap-free recording possible but shorter than asked. --force records it and then stops (stop_reason ring_full, ok, with a warning); the rest of the requested time is not recorded. Both errors carry the numbers in error.plan.
5. While running, every statistics line has backlog_mb; while the backlog grows, also backlog_growth_mb_s, ring_full_in_s and gapfree_projected_s. The result has unread_at_stop_mb (acquired, never read).

Examples here: 12 channels at 50 MSa/s (250 MB/s, R = 204): gap-free 8.88 s vs stream records of 2.03 s with 0.64 s gaps, so NEEDS_FORCE; forced, the run stopped after 8.84 s (8.30 s with R = 202 in an earlier run), every period present up to the last sample. 12 channels at 100 MSa/s (500 MB/s, R = 264): 1.13 s vs 1.02 s records with 1.09 s gaps, still NEEDS_FORCE; with a readout below ~250 MB/s stream would win. 200 MB/s read at 190 MB/s: 10 MB/s backlog growth, the guard after ~50 s, 47.8 s of gap-free signal.

The backlog is estimated as D x (time since the head) - bytes received, checked on every USB transfer, against a baseline that follows the scope / PC clock difference: it may rise by at most 200 ppm x D per second, and only while the host is caught up (estimate below 16 MiB), so the estimate stays exact over hours and a real backlog is never absorbed. A guard stop while nothing arrives at all is reported as a stalled scope (TIMEOUT_DATA); a forced log that was planned to end at the guard reports ring_full, any other guard stop HOST_TOO_SLOW.

log or stream?

Both record untriggered, but they are not the same mode of the scope. log is one endless run (loop mode); stream is a series of records, each programmed with a fixed length that the scope stops at.

The restart (~0.18 s) is the same for both; the difference is what has to be read out before it. Use log for continuous recording; stream for different analog and digital rates, decoding the record, or above the readout limit when log says STREAM_BETTER or the whole time matters more than one long gap-free piece.

python -m openm454 bench --ram 8GB --disk 50GB --dir logs
python -m openm454 log --plan -c A0,D0-D7 --rate 1M --duration 10min
python -m openm454 log -c A0,D0-D7 --rate 1M --duration 10min                   # kept in RAM, one .npz at the end
python -m openm454 log -c A0,D0-D7 --rate 1M --duration forever                 # RAM ring; end it with: python -m openm454 stop
python -m openm454 stream -c D0-D7 --rate 100M --duration 60s --every 1s

Every command prints one JSON document (watch, stream and log: one per line, then the result), whether or not --json is given: measurements, applied settings, file paths. Samples are never printed (unless --dump N asks for a decimated series); captures are saved as .npz (raw codes + scaling) and .png. In the plot the analog legend is one row above the plot area (one entry per channel, e.g. A1 1V/div Vpp 2.86V f 1kHz) with the title above it, digital lane labels sit in the right margin, and only the trigger markers (dashed line at t = 0, dotted analog trigger level) are drawn over the waveform; decoded frames are not drawn, they are in the JSON and <capture>_decode.json. python -m openm454 <command> --help lists every option. Python API: openm454.scope.Scope, openm454.capture.load_capture, Capture.decode(), openm454.plot.render(cap, window=None, title=None, summary=None, channels=None) (returns the figure; the caller closes it), openm454.firmware.status() / save_dswave().

Protocol decoders (--decode SPEC, repeatable, on capture, watch, stream and analyze): uart:rx=D4,baud=115200|auto,bits=8,parity=none,stop=1,invert=0, spi:clk=..,mosi=..,miso=..,cs=..,mode=0..3, i2c:scl=..,sda=.., can:rx=..,bitrate=500k|auto,invert=0. They report every frame with trigger-relative times and its errors (parity/framing, partial words, NACKs, CRC/stuff/form errors), and remove the comparator chatter of slow edges. Output format: docs/schemas/decode.json. Decoding needs about 10 samples per bit or clock (I2C: 9 per SCL period, no errors; 5-6 per period still decode with well under 1 % errors). They have been checked against real buses (SPI up to 18 MHz, RS232, CAN up to 1 Mbit/s, I2C up to 400 kHz) generated by a microcontroller, including deep captures and decoded streams.

CLI reference

pip install openm454 installs the openm454 command (python -m openm454 ... is the same) and the openm454-mcp server. Global options, before the subcommand: --max-wall T (watchdog: stop and release the device after T or the budget computed from the command, whichever is longer; it never shortens the computed budget), --verbose (wire log on stderr), --json (accepted, output is always JSON). python -m openm454 <command> --help lists every option.

capture, watch and stream share these options:

• channels / vertical: -c A0-A3,D0-D7 (also analog, digital), --vdiv 1V, --offset 0V, --coupling dc|ac, --bw20 A1,A2, --invert A2, --probe 10x|A2=10x, --threshold 1.4V. Per-channel values are written A0=500mV,A2=2V.
• timing: --timebase (window = 10 divisions), --duration, --samples, --rate.
• trigger: --trig SRC:EDGE[:LEVEL] (e.g. A1:rise:1.4V, D0:fall, D0:pulse>:200us), --trig-mode auto|normal|force (default auto), --trig-pos 50%, --auto-after 0.3s, --timeout 5s (normal-mode wait), --holdoff, --sensitivity 0..15 (default 5), --trig-retries N (default 4; digital edge triggers, see Digital edge triggers). stream accepts the trigger options and ignores them: it always records untriggered.
• output: --measure basic|full|none (see Measurements), --edges N, --dump N, --window t0:t1, --no-plot, --no-save, --out, --tag, --awg 'sine,10k,0V,2V[,duty%]' (set the generator first), --decode SPEC (repeatable), --decode-limit (default 50 frames).

log takes the vertical options too; its timing is --rate and --duration (see the recording sections above). A value that starts with - needs the = form on Python 3.10-3.12 (--window=-1ms:1ms, --offset=-1.5V; per channel A1=-1.5V is fine as it is); the MCP server does this itself. --trig without a level on an analog channel uses that channel's --offset as the level. --rate snaps down to a rate the scope has (at most 400M on 1 analog channel, 200M on 2, 100M on 3-4; digital only: 400M on up to 2 channels, 200M on up to 4, 100M on 5-8). capture keeps only the newest 200 captures with its --tag (default cap) in its output folder (captures/ by default, or --out): older files named the way openm454 names them (<yyyymmdd-hhmmss-mmm>_<tag>.npz, .png, _decode.json) with that tag are deleted. Recordings with other tags (log, stream, watch_last) and any other file in the folder are never touched.

Exit codes: 0 ok, 1 other (a failed self-test, an internal error, interrupted), 2 bad arguments, 3 scope not found, 4 busy, 5 no trigger, 6 device / protocol / watchdog failure, 7 refused by the safety rules (see Errors).

watch

watch repeats capture and prints one JSON line per capture.

• --every T (default 0.5s) is the time from the start of one capture to the start of the next. If a capture takes longer, the next one starts at once, with no catch-up burst.
• t_s of a line is when that capture started, in seconds since the watch began (--every 0.5s gives 0, 0.5, 1.0, ...). The final line has elapsed_s.
• --count N and --for T end the watch; with both, whichever comes first. No capture starts at or after --for: --every 0.5s --for 2s gives 4 captures. With neither, 20 captures. openm454 stop ends the watch after the current capture, or within 0.25 s while it waits between captures.
• --fields CH.KEY,...: KEY is any measurement name; only the named channels are measured. The default fields are vpp_v, vmean_v, freq_hz, duty_pct and state of every channel.
• --stop-when EXPR: a Python expression with one dict per channel, e.g. "A1['vpp_v'] < 1"; an expression that raises counts as false. --fields or --stop-when together with --measure none is refused (exit 2).
• --window and --edges apply; --dump, --no-plot, --no-save and --tag have no effect. --out is only used by --save-last, which writes <out>/<date>-<time>_watch_last.npz.
• A line adds trigger_check when it is not ok / unchecked, and trigger_misfires when above 0.

Measurements (--measure)

--measure basic|full|none applies to capture, watch, the final result of stream and analyze.

• basic (default). Analog channels: samples, vmin_v, vmax_v, vpp_v, vmean_v, vrms_v, vac_rms_v, vtop_v, vbase_v, vamp_v, freq_hz, period_s, duty_pct, edges_rise, edges_fall, periods_in_window, pwidth_s, nwidth_s, rise_s, fall_s, overshoot_pct, undershoot_pct, clip_lo, clip_hi, and when they occur freq_method, period_irregular, noise_like. Digital channels: samples, state, high_frac, edges, freq_hz, period_s, duty_pct, and when they occur glitch_edges and runts. On analog channels, frequency and duty come from falling edges when the window holds at least two falls but fewer than two rises (digital channels need two rising edges).
• full: basic plus, per channel, statistics over every whole cycle (rising edge to rising edge) in the window: cycles, period_min_s, period_max_s, period_std_s, freq_min_hz, freq_max_hz, pwidth_min_s, pwidth_max_s, nwidth_min_s, nwidth_max_s, duty_min_pct, duty_max_pct (digital channels also get the mean widths pwidth_s, nwidth_s), and cycles_capped: true when the 100k-edge limit was reached (the statistics then cover the first 100k edges). Not given for channels with fewer than 2 rising edges or for noise. Analog channels also get range_lo_v and range_hi_v, the voltages the screen can show at the applied V/div and offset; readings at the ends are clipped. Edge lists (edges_rise_s, edges_fall_s) appear only with --edges N. Deep records take a second measuring pass.
• none: no analog / digital blocks (hints are still given). The MCP detail argument (summary | full) is a different thing: it chooses a compact summary or the CLI's whole document.

Digital edge triggers

The digital inputs have no hysteresis. A slow edge (for example the ~8 us rising edge of the calibration square) chatters at the threshold for tens of ns, so a falling-edge trigger can fire inside the rising edge, and the reverse, while the hardware still reports a valid trigger. After every capture or watch triggered on a digital channel with a rise or fall edge (normal or auto mode), openm454 therefore checks the settled level of the trigger line just before and after t = 0 (a level is settled when it lasts at least 200 ns and at least 3 samples, whichever is longer; shorter runs are chatter). It only calls a capture a misfire when the unsettled stretch between the two settled levels is as short as chatter (at most 200 ns; measured chatter lasts up to about 120 ns); a longer stretch is real data and gives inconclusive. So fast data right at the trigger (faster than about 10 Mbit/s, e.g. a short 1-0 pair of bits between a long low and a long high) can be taken for chatter and re-armed, and an edge so slow that its chatter runs last 200 ns or more passes as ok. A capture that fired on the opposite edge is discarded and the scope re-armed, up to --trig-retries N times (default 4; 0 = only check; MCP trig_retries). Re-arms share the original trigger wait (--timeout in normal mode, --auto-after in auto mode, plus --holdoff); only time spent waiting for a trigger counts. Analog triggers, any and pulse triggers, --trig-mode force and the untriggered fallback of auto mode are not checked.

The capture's acq block reports it:

• trigger_check: ok (the first capture shows the requested edge); misfire_retried (misfires were discarded, the returned capture is correct); mismatch (the returned capture is itself a misfire: t = 0 is NOT the requested edge, with a warning and a hint); inconclusive (the edge could not be confirmed, e.g. no pre-trigger samples at --trig-pos 0, or a glitch or clock faster than 200 ns; never re-armed); unchecked (not a digital rise/fall trigger, or untriggered). MCP summaries leave out unchecked.
• trigger_misfires: the number of misfired captures discarded and re-armed (left out of MCP summaries when 0).
• attempts lists every run: triggered, forced (untriggered), misfire (fired on the wrong edge), no-trigger. acq.triggered keeps its meaning: the hardware trigger flag of the returned capture.
• Warnings and hints: mismatch gives a warning and a hint; misfire_retried also adds a hint with the number of discarded misfires (e.g. "D5 falling-edge trigger: N misfire(s) inside a rising edge were discarded and re-armed").

Measured on the M454 (1 kHz calibration square, normal mode): before the check, a D5 falling trigger fired inside the rising edge in 13 of 40 captures at 100 MSa/s, D0 falling in 5 of 40, D5 rising never. With the check, 60 captures each: no wrong edge returned at 50, 80 or 100 MSa/s (D5 alone, D5 with A1, all 8 digital channels; 5-25 misfires re-armed per set). At 400 MSa/s on D5 alone about half of all tries misfire, and 2 of 60 captures misfired on all 5 tries: both were returned as mismatch with the warning, not as a success. For a reliable digital trigger use the edge that is fast on the signal (a falling trigger misfires on a slow rise), or trigger on an analog channel, whose trigger has a noise margin.

Vertical limits

• --offset is the voltage at the centre of the screen. The allowed range is about -4.32 .. +4.36 divisions (-127 .. +128 offset codes), where a division is the hardware V/div (the requested value rounded up to a 1-2-5 step, 1 mV .. 5 V at the probe input). At 1x:

  V/div	allowed offset	V/div	allowed offset
  1 mV	-4.323 .. +4.357 mV	100 mV	-0.4323 .. +0.4357 V
  2 mV	-8.647 .. +8.715 mV	200 mV	-0.8647 .. +0.8715 V
  5 mV	-21.62 .. +21.79 mV	500 mV	-2.162 .. +2.179 V
  10 mV	-43.23 .. +43.57 mV	1 V	-4.323 .. +4.357 V
  20 mV	-86.47 .. +87.15 mV	2 V	-8.647 .. +8.715 V
  50 mV	-216.2 .. +217.9 mV	5 V	-21.62 .. +21.79 V

  Every number scales with the probe factor (at 10x: 10 mV .. 50 V/div at the tip, limits x10).

• A larger offset is clamped with a warning, e.g. A1: offset 2.5 V out of range at 0.5 V/div (allowed -2.162 .. 2.179 V, about +-4.3 div); clamped to 2.179 V; the value used is in applied.offset_v. watch lines show no warnings.

• The trigger level must be on the screen (codes 10..245 around the offset), otherwise TRIGGER_OUT_OF_RANGE (normal and auto mode; with --trig-mode force an off-screen level is clamped without an error). Without a level, an analog trigger uses --offset.

• Clipping gives a hint (A1 clips at the top: increase --vdiv or lower --offset); the front end needs ~0.2-0.3 s to recover after an overdrive.

Self-test

python -m openm454 selftest checks the scope itself. The M454 has a 1 kHz calibration output (a ~3 V square wave, the probe-compensation signal) and the signal generator output, but no internal loopback: the inputs are not connected to either inside the scope. The standard fixture is therefore the scope's own two outputs cabled to its inputs, with no other instrument. The checks fall in two groups.

Wiring-free checks run on any wiring. They only read the inputs and never touch the generator (about 4-6 s):

• V01 device info
• V12 a forced capture of all 12 channels (A0-A3, D0-D7) at 100 MSa/s
• V13 the top rates and fold modes: 1 analog channel at 400 MSa/s, 2 analog at 200 MSa/s, 2 digital at 400 MSa/s, 4 digital at 200 MSa/s
• V11 capture latency

Signal checks (V00, V02-V10: calibration square, V/div, sample-rate consistency, digital channels and threshold, trigger position, digital trigger, no-trigger timeout, generator loopback) need the standard fixture below and are reported as skipped, with the reason, when the wiring is not declared as that fixture. Their tolerances are, for instance, 1 kHz +-2 %, top 2.8 V +-12 %, base 0 V +-0.15 V, duty 50 +-3 %.

Declare the wiring once, in state/fixture.json of the data folder, or for a single run with --fixture standard:

python -c "from openm454 import fixture; fixture.save('standard', 'standard')"
python -m openm454 selftest                      # uses the recorded wiring
python -m openm454 selftest --fixture standard   # the fixture is wired right now (this run only)

fixture.save(analog='standard', digital='standard', note=None) writes {"analog": ..., "digital": ..., "since": ...}; "standard" for a group of leads means the standard fixture (any other value means something else is connected and the signal checks on those leads are skipped). A missing file means unknown wiring, so a fresh install runs only the wiring-free checks; so does a file that cannot be read (the result's note says why; a UTF-8 byte-order mark is fine). Update the file whenever the wiring changes.

The generator is used only with the analog fixture and without --no-awg: V09 parks it at 0 V for the no-trigger timeout, and V10 drives a 2 Vpp sine and a 0-3 V square into A0, then turns it off (also when one of its captures fails). --no-awg (and the MCP default) leaves the generator untouched; V10 is then left out of the results, and V09 is skipped when the last awg setting left the generator running (run awg off first to include it). --quick runs fewer repetitions (about half the time). The result lists every check with measured and expected; the exit code is 1 when a gating check failed.

Which command for what

Monitoring with log

log prints one statistics line per --every of signal time (default 1 s), so a long run is a table of vmin_v / vmax_v / vmean_v per interval: dips and spikes show in vmin_v / vmax_v even when the mean hides them. A 3.3 V rail for 8 hours, one line per ~10 s, nothing saved (--storage none):

python -m openm454 log -c A1 --vdiv 0.5V --offset 2.1V --rate 10k --duration 8h --every 10s --storage none

The window is ~0.09..4.09 V and one ADC step is ~17 mV (pick V/div as described under Limits and safety). Without --storage none every sample is kept in RAM (the default) and written at the end: 72 MB for 2 h at 10 kSa/s, 288 MB for 8 h. Ripple on a rail: AC coupling and a small V/div, which loses the DC level:

python -m openm454 log -c A1 --coupling ac --vdiv 20mV --rate 1M --duration 1h --every 1s --storage none

Check a plan first with log --plan (same options, no device). Over MCP: log_plan, then record_start with mode: "log" (it takes vdiv, offset and coupling too), record_status (the last lines) and record_stop. A recording started over MCP is stopped when the MCP client exits, so an unattended multi-hour log needs the Claude session to stay open, or the CLI run in the background (openm454 log ..., ended with openm454 stop).

How fast can log refresh?

--every can be as small as one 64-sample block, and lines are cut at exact signal-time boundaries (--every rounded to whole 64-sample blocks, see plan.every_s), not at USB transfers. Measured on the M454 at 150 MB/s (A0-A1 + D0-D7 at 50 MSa/s, --storage none, 10 s each):

Printing 1000 lines/s while recording 150 MB/s is therefore not the limit (the lines carry statistics, not the samples). The data reaches the PC in USB transfers of ~1 MB (about 7 ms at 150 MB/s), so lines arrive in bursts and the earliest reaction is ~10 ms after the signal. The real limit is the reader: an agent reading JSON lines or polling MCP record_status reacts in seconds at best. To react within milliseconds, run your own code next to the data with a hook.

Live lines of log and stream

log: when this channel set's readout is measured first (above 150 MB/s and not yet known, or --remeasure), a probe line; then a line with "started": true, the plan and out_stem; then one line per --every, then the result. With --hook, event lines ({"event": ...}, no i) are interleaved as they happen; see Live processing hook. Lines are cut at whole 64-sample blocks, so a line covers --every rounded to a multiple of 64 / rate (at 10 kSa/s --every 10s gives 9.9968 s); plan.every_s shows the value used. A live line has:

stream: lines have i, t_signal_s, t_wall_s (seconds since this record's start; it restarts at every record of a chained or forever stream, while t_signal_s runs over the whole stream), A0..A3 with vmin_v, vmax_v, vmean_v and n (samples in this line), and D0..D7 with high_frac, edges. A line is emitted once signal time passes the next --every boundary, so it is cut at USB event boundaries. A single record ends with final: true; a chained stream adds record: k to every line and record_end: true at the end of a record. There are no backlog fields. The final result has the full measurements (--measure, --window, --edges, --dump), files, hints and warnings; chained streams add records, record_s, signal_s, seams_s, seam_max_s, stopped_by.

Live processing hook (log --hook)

A hook is a Python file of your own that sees every sample of a gap-free log recording as it arrives and can react at once: a threshold alarm, an edge counter, a trigger for other equipment, an early stop. It is the way for an agent to act on real-time data faster than it can read lines (see Monitoring with log).

python -m openm454 log -c A1,D0 --rate 1M --duration forever --storage none --hook myhook.py

The file is loaded before the scope is opened. Loading imports the file, so its top-level code runs. A missing file, a syntax or import error, a sys.exit(), or a missing process() gives INVALID_ARG (exit 2) and nothing is recorded. With --plan the file is loaded and checked the same way; no hook function is called and the scope is not touched. The file runs as a registered module (dataclasses and the like work), its folder is added at the end of the import path (helper modules next to it import), and print() in it goes to stderr, so stdout stays one JSON document per line.

The file

def setup(info):      # optional, once before the first sample
    ...

def process(block):   # required: called for every piece of the recording, in order
    return None       # or a dict: an event

def finish(summary):  # optional, once at the end; a returned dict goes into the final result
    ...

info has rate_hz, analog, digital, scale (per analog channel: volts_per_code, zero_code, probe, vdiv_v), every_s, duration_s, storage, block_samples, piece_s (the planned piece size in seconds). summary of finish() has samples, duration_s, stop_reason, calls, events.

Every sample is passed exactly once, in order, without gaps, in pieces of about 10 ms of signal (shorter when --every is shorter, since a piece never spans more than one line; longer when a USB transfer holds more, e.g. ~20 ms roll packets at slow rates; info['piece_s'] is the planned size). Lines, statistics and storage are unchanged by the hook.

The block

The arrays are the recording's own and read-only: an in-place edit (c -= 128) raises and ends the run with HOOK_ERROR, the stored data unchanged. Compute on a copy (c.astype(np.int16) - 128). Pieces vary in length (n): carry what you need across calls (the last sample, a running state) or an edge or crossing at a piece boundary is missed.

Events and stopping

process() returns None or a dict. A dict is an event and is printed at once as {"event": {...}, "k": n, "t_signal_s": ..., "lag_s": ..., "t_wall_s": ...}: t_signal_s is the end of the piece, lag_s how far behind live the hook was (seconds of newer signal the scope had already taken), k the event number from 0, t_wall_s PC seconds since the run started. Put "stop": true in the dict to end the recording cleanly (acq.stop_reason hook, ok true): the recording ends with that piece, and nothing after it is kept or counted (in a RAM ring the history before the event stays). Any other return type is a HOOK_ERROR.

The dict is printed as JSON: numpy scalars and arrays in it become plain numbers and lists, NaN and infinity become null; anything else JSON cannot hold (a set, an object) is a HOOK_ERROR. The same holds for the dict finish() returns.

Live lines get hook_ms (time spent in the hook since the last line). The final result gets a hook block: file, calls, events, seconds, load (hook seconds per signal second), lag_max_s, stopped_run, last_event and result (the dict finish() returned, if any). A load above 1 or a lag_max_s above 1 s adds a warning that the hook fell behind.

Hook errors

Any exception raised by hook code, sys.exit() and KeyboardInterrupt included, is a HOOK_ERROR. An exception in process() ends the run with HOOK_ERROR (exit 2; error.trace has the traceback tail; acq.stop_reason is hook_error). The data recorded up to and including that piece is kept and saved, and finish() is not called. An exception in finish() also gives HOOK_ERROR with error.trace, but the recording is complete and acq.stop_reason stays what ended it. An exception in setup() gives HOOK_ERROR (the message names the exception, there is no error.trace) before anything is recorded. When the run ends with a device error or Ctrl+C, finish() is still called and the error result carries the hook block.

Hook performance

The hook runs in the thread that decodes the data, so it has to keep up. If it falls behind, the queue grows and the host-speed guard stops the run with HOST_TOO_SLOW at 768 MiB (~805 MB) queued (data kept). Measured at 150 MB/s: a light hook (count D0 edges with bits(), min/max of A1 codes, an event every 100 ms, stop at 2 s) made 144 calls in 2 s with load 0.022, lag_s median 0.2 ms and max 7 ms, and stopped the run itself at 2.0 s. A heavy hook (volts() of both analog channels plus an FFT per piece) had load 3.16, lag up to 2.2 s and a warning. So: use numpy on codes() / packed() instead of volts() / bits() for every sample, keep the work per piece small, accumulate across pieces in module-level state, and lower --rate or use fewer channels if needed.

Hook safety

The hook must not talk to the scope: it is busy recording, so no openm454 commands from the hook. Files, sockets, serial ports and subprocesses are fine. A hook is your own code running with your rights: only run hook files you wrote or trust.

Hook over MCP and from Python

record_start takes hook (the absolute path of the file; a relative path is refused because the recording runs in the MCP server's folder) with mode: "log"; with mode: "stream" it is INVALID_ARG. The hook's imports must be installed in the Python that runs the MCP server. record_status then adds events (the last lines event lines, default 3, at most 20) and events_seen (event lines in the last 256 KiB of the output; the total is hook.events in the result of record_stop). From Python: logger.run(..., emit=print_or_collect, hook=openm454.hook.Hook(path)); the event lines go to emit.

Example: power-rail dip alarm

A1 on a 3.3 V rail; the hook reports the first sample below 3.0 V in each piece, comparing raw codes against a precomputed integer limit, so no volts array is built per piece:

import numpy as np
LIMIT_V = 3.0

def setup(info):
    s = info['scale']['A1']
    global LIMIT_CODE
    LIMIT_CODE = int(np.ceil(s['zero_code'] + LIMIT_V / (s['volts_per_code'] * s['probe'])))

def process(block):
    c = block.codes('A1')
    i = np.flatnonzero(c < LIMIT_CODE)
    if i.size:
        return {'dip_at_s': round(block.t0_s + i[0] / block.rate_hz, 6), 'min_v': round(float(block.volts('A1').min()), 4)}

python -m openm454 log -c A1 --vdiv 0.5V --offset 2.1V --rate 1M --duration forever --storage none --hook dip.py

Run on the M454 with A1 on the calibration square, which is below 3.0 V most of the time: one event per ~20 ms piece at 1 MSa/s, load 0.005. The light and heavy hooks under Hook performance were measured on the M454 too.

Limits and safety

From the vendor datasheet:

• The probe ground is the PC's ground (through USB, and through the PC's mains earth if it has one). Connect it only to a point at the same potential, never to a "hot" ground or any point referenced to the mains.
• The on-screen range is not the safe range. At 1x the screen reaches about +-41.7 V (5 V/div with the offset at its limit), and at 10x about +-417 V, but the BNC limit is +-200 V peak and the probe's own rating applies.
• Running past the screen saturates the ADC; the data of that part is clipped.

Choosing V/div and offset for a DC rail with DC coupling: the screen shows 8 divisions, the farthest reachable edge is +8.34 divisions, so use V/div >= V_rail / 8.34 and set the offset so the rail is on screen: with the default offset 0 V a positive rail above ~4 x V/div is off screen (at 1 V/div the window is -4.02..+3.98 V).

MCP server

openm454-mcp (or python -m openm454.mcp_server) serves the scope over the Model Context Protocol (stdio, no extra dependencies; it is installed with pip install openm454):

claude mcp add -s user openm454 openm454-mcp

This form works in every shell (PowerShell, cmd, bash). -s user makes the server available in every folder (without it, only in the folder where you ran the command). If pip's Scripts folder is not on PATH, give the full path to openm454-mcp. DSWave is found by itself (search order above), or record it once with python -m openm454 setup --dswave <folder>. Check with claude mcp list. Other clients: {"mcpServers": {"openm454": {"command": "openm454-mcp"}}}.

An environment variable (OPENM454_DSWAVE, OPENM454_HOME) reaches the server only through the environment Claude starts it with: give it when registering (-e). Claude starts the server in the background, so a variable set later in a shell never gets there. Put the server name before -e (-e takes several values and would swallow a name after it), then --, then the command:

claude mcp add -s user openm454 -e "OPENM454_DSWAVE=<folder>" -- openm454-mcp

In Windows PowerShell, an npm-installed claude runs through a .ps1 wrapper that drops a bare --, which gives "missing required argument". Write the separator in quotes ('--'), call claude.cmd, use "--env=OPENM454_DSWAVE=<folder>" with no separator, or run the command in cmd. Restart Claude afterwards. Each tool call runs one CLI command in a child process, so the scope is opened and released per call, a hang never blocks the server, and every rule of the CLI applies.

11 tools:

Results are summaries; on capture, analyze and selftest, detail: "full" returns the CLI's whole document. The samples stay in the .npz. Every device call has a hard time limit, after which an error (TIMEOUT_WATCHDOG) is returned and the CLI's own watchdog stops the scope: capture and selftest take max_wall_s (default for capture: 180 s plus the trigger timeout and the record length; for selftest: 1260 s, above the CLI's own 1200 s watchdog); status, awg and release are limited to 60 s.

selftest runs selftest --quick (--quick dropped when quick is false), with --no-awg unless awg is true and --fixture standard when fixture is standard, so by default it never touches the generator and only runs the signal checks if the wiring is declared. Pass fixture: "standard" only after the user has wired and confirmed the standard fixture, and awg: true only when the generator output goes to A0 and nothing else. The summary has passed_checks, failed_checks (measured vs expected), skipped_checks (grouped by reason) and counts that match those lists (checks not run after a device failure count as skipped), plus fixture, note and seconds. A failed run returns the error SELFTEST_FAILED, whose hint tells a failed signal check (usually the wiring) from a failed wiring-free one (the scope or its USB link); a device failure returns that device error (with stopped_at).

CLI commands with no MCP tool:

An agent that has a shell can use the CLI for all of these.

Errors

Every error has a code, a message and a hint, so the next step is clear:

Layout

• openm454/ driver: USB transport with a safety whitelist, device bring-up (FPGA, security handshake, calibration), register compiler (mso.py, reproduces DSWave's register programs exactly), stream decoder (packet.py), acquisition engine, loop-mode logger (logger.py), live processing hook for log --hook (hook.py), measurements, plots, AWG, CLI, self-test, protocol decoders (decode.py), synthetic bus traffic for tests (protosim.py) and the MCP server (mcp_server.py).
• tests/ offline tests: golden register and AWG programs, the stream decoder, the acquisition engine against a simulated device, measurements, protocol decoders on synthetic traffic, the MCP server, safety rules (python -m pytest tests).
• docs/PROTOCOL.md the recovered USB protocol.
• .github/workflows/: CI (the offline tests on Windows and Linux, Python 3.10-3.14, with OPENM454_NO_USB=1, which makes the driver refuse to open USB; a check that tracked files carry no local paths; a package build) on pushes to main and on pull requests; a tag vX.Y.Z equal to openm454.__version__ tests, builds and publishes to PyPI (trusted publishing) and creates a GitHub release.

What is NOT here

The vendor application, its embedded FPGA bitstream and firmware images, decompiled vendor code, raw USB traces (they contain the device's security key), the reverse-engineering workspace and the vendor datasheet are kept out of this repository.

Safety

The driver never sends the firmware-upgrade request (bRequest 0xB3), never writes the device EEPROM, and configures the FPGA only when the device reports it unconfigured or when openm454 recover resets a hung scope (always from the verified bitstream, as the first write of the session).

License

GPL-3.0 (see LICENSE). DSView (GPLv3) by DreamSourceLab was used as a protocol reference. DSWave and DSCope are trademarks of DreamSourceLab; this project is not affiliated with DreamSourceLab.