All servers / Developer / Denpex MCP server
Denpex MCP server
Denpex MCP server, diagnose GPU/ML training failures (CUDA OOM, NCCL timeouts, Xid faults, NaN loss) from Claude Desktop, Cursor, or Claude Code, with optional
Tested by ConnectATDScore 85 out of 100. Tested 2026-11-19 on package version 0.2.46. Package: denpex-mcp · source
Score breakdown
| Runs | 18 / 20 | |
|---|---|---|
| Tool descriptions | 19 / 20 | |
| Package safety | 20 / 20 | |
| Maintenance | 20 / 20 | |
| Adoption | 8 / 20 |
What we saw
- Starts and answers the MCP handshake: yes (4.0 seconds including download)
- Tools offered: 18
- Last published: 2026-11-18
- Weekly downloads (npm): 790
- License: MIT · Direct dependencies: 2
Want this server set up for your team?
ConnectATD builds AI automation and wires up MCP servers like this one. Tell us what you are trying to do.
Talk to ConnectATDTools and how they are described
Agents choose tools from these descriptions. Under each name you can see how many of its parameters are explained.
diagnose_crash12 of 12 parameters describedDiagnose a GPU/ML training crash (CUDA OOM, NCCL timeout, Xid fault, NaN loss, checkpoint corruption, etc.) and get the root cause plus an escalating fix (primary → fallback → emergency). Provide ONE of: `logs` (paste), `file` (path to a log), or `dir` (folder to scan for the newest *.err/*.log). Runs the canonical deterministic engine locally by default (unlimited free local runs), and escalates to the Denpex cloud when DENPEX_API_KEY is configured or cloud: true is passed. Or run fully offline with DENPEX_LOCAL=1.
inspect_gpu_artifact1 of 2 parameters describedRead-only local inspection of PyTorch Flight Recorder JSON rank dumps, CUDA memory snapshot JSON, Nsight SQLite, checkpoint candidates or aligned rank timing JSON. Requires Python 3; uses no network and never executes pickle or changes checkpoints. Pass the resulting JSON to diagnose_crash as logs to review or save it using your normal Denpex configuration.
lookup_failure2 of 2 parameters describedSearch Denpex's bundled 721-class failure encyclopedia by error code, keyword, or symptom (e.g. 'Xid 79', 'NCCL timeout', 'loss is NaN', 'CUDA out of memory'). Returns the most relevant known failures with their fixes. Works fully offline and costs no quota. Use `diagnose_crash` instead when you have a full crash log.
analyze_cascade3 of 3 parameters describedAnalyze a multi-rank distributed-training failure: find the rank that failed FIRST (the others usually just time out waiting on it) and route any GPU Xid faults to the correct node action (reset / reboot / RMA). Runs offline and instantly, provide the gathered per-rank logs via `logs`, `file`, or `dir`.
check_quotano parametersShow your Denpex plan, allowance tier, and diagnoses left today.
ask_followup4 of 4 parameters describedAsk the Denpex diagnostician a follow-up question about a crash, clarify the fix, explore an alternative, or dig deeper. Pass the original `logs` and your `question`; pass back the `conversationId` from a previous reply to continue the thread. Requires the Team plan or an active trial.
check_versions3 of 3 parameters describedScan logs (or an environment dump like `pip list` / `nvidia-smi`) for CUDA / PyTorch / cuDNN / NCCL / driver versions that are known to be incompatible or buggy together, and get the recommended pin. Requires the Team plan or an active trial.
list_incidentsno parametersList the GPU/training incidents Denpex's always-on monitor caught and auto-escalated on your fleet, failure type, cluster, escalation level, and the fix it applied. Requires the Scale plan or higher.
recent_diagnoses1 of 1 parameter describedShow your most recent Denpex diagnoses, job name, failure type, confidence, and the fix, so you can recall what a prior crash was and how it was resolved.
submit_fix_outcome4 of 4 parameters describedReport back whether a Denpex fix actually resolved the failure. Call this after you apply a fix from `diagnose_crash` and observe the result, it is the ONLY way the engine learns, and it directly improves the next diagnosis of this failure class for you and everyone else. Pass the `diagnosisId` printed at the bottom of the diagnosis. If the fix did not work, say what actually happened in `note` (and `newFix` if you found the real one), a specific negative is worth more than a vague positive.
fleet_diagnostic2 of 2 parameters describedAnalyze output from a GPU fleet tool (gpud, gpu-burn/DCGM, nvbandwidth/perftest, py-spy, nvitop/nvtop, HTA traces, genai-perf/vLLM bench, hwloc/lstopo, LBNL NHC, SuperBench, dynolog, memray/scalene, hf-mem/torchsnapshot, nccl-tests). Leave toolId as "auto" to detect the tool from the output. Needs a Denpex API key: Team and higher plans cover all 14 tools, Free covers 05-nvitop-nvtop, 08-hwloc-lstopo and 13-hf-mem-snapshot. For crash logs use diagnose_crash instead. Use the local Denpex CLI for files that must stay on the host. The report is advisory.
launch_gate3 of 6 parameters describedEvaluate the EXACT allocation a scheduler is about to hand out, these nodes, these GPU indices, against what Denpex knows about them: quarantined or repeat-offender cards, split driver versions inside one collective, software combinations that have crashed and never completed, degraded fabric links, an incomplete distributed checkpoint, storage about to hit ENOSPC. Returns GO / GO_WITH_WARNING / REQUIRE_APPROVAL / BLOCK with the evidence for each. Advisory unless `enforce` is true. IMPORTANT: a check Denpex could not evaluate is reported as unknown, never as a pass, read `evidenceCoverage` before treating a GO as an all-clear.
preflight_check3 of 3 parameters describedAnswer 'is this cluster safe to launch on?' BEFORE you burn GPU-hours on a job that will die in the first ten minutes. Feed it the health artifacts from the nodes you are about to use, `nvidia-smi -q`, `ibstat`, `nvidia-smi topo -m`, `dmesg -T`, and it returns the critical blockers (dead links, missing GPUs, non-NVLink topology, driver mismatch) that would take the run down. Scale+ feature.
compare_runs2 of 2 parameters described'It worked yesterday.' Give it the logs from the run that worked and the run that failed, and it isolates what actually changed between them instead of re-diagnosing the failure from scratch. Use this whenever a previously-green job starts failing.
diff_environments3 of 3 parameters describedDiff the SOFTWARE environment between a good run and a bad one, torch / CUDA / NCCL / driver / flash-attn versions, env vars, and flag combinations known to be broken. Feed it `pip freeze`, `nvidia-smi`, `env` output (or the whole job log; it extracts what it needs). Use this when `compare_runs` points at a version change, or when a job breaks after an image rebuild.
validate_checkpoint4 of 4 parameters describedInspect directory listings, checksum output and load errors for signs of checkpoint corruption. Supplied metadata does not verify full shard coverage, compatible optimizer state or successful workload recovery. Validate actual loading and intended progress before resuming.
node_healthno parametersScore the health of every node Denpex is monitoring in your fleet, worst first. Use it to answer 'which nodes should I exclude from this job?' before scheduling, or 'what else is degrading?' after an incident. Scale+ feature; requires the Denpex agent to be reporting.
predict_failuresno parametersList the predictive risk scores Denpex has computed for your fleet, nodes whose counters are trending toward a failure that has not happened yet (rising correctable ECC, row-remap pressure, PCIe replays, thermal drift). Use it to drain a node on your schedule instead of the cluster's. Scale+ feature.
Tested in a sandbox with an empty environment: no API keys, a throwaway folder, nothing from a real account. A passing test shows the server runs and responds. It is not a security audit.