Locate the local NVIDIA Nsight Graphics ngfx-replay executable and report its version and MCP cache status.

Read capture identity, application/GPU/API/version metadata, portability flags, and replay incompatibilities without replaying the frame. Environment values are never returned.

Build a bounded static analysis from Nsight metadata, API event names, and object identities: stream shape, top calls, thread distribution, resource population, compatibility/security findings, and explicit limitations.

Page through the exported API function stream. Filter by function-name regex, inferred category, thread, or event-index range. Results contain names and indices, not call parameters or timings.

Page through captured object identities and filter by UID, API, type, object-name substring, creator-function substring, or nonzero access flags.

Read captured Nsight/application log messages with optional error-only mode, text search, and pagination. Credential-shaped values are redacted.

Extract the final-present screenshot embedded in a capture. Without output_path it returns an MCP image using a temporary file. With output_path it also saves the image; existing files require overwrite=true.

Run Nsight GPU Trace on one hidden replay pass and return the Summary-tab range, Top-Level Throughput, Active Threads Per Warp, and SM Warp Occupancy data. This executes instrumented GPU work and can take several minutes for large captures.

Run a feature-rich Nsight GPU Trace replay, retain its report and auto-export tables in a report session, and return a report_id. Query that id with the metric, range, trace-information, and shader-occupancy tools. Collection executes instrumented GPU work and can take several minutes.

Open an existing Nsight GPU Trace auto-export directory and register it as a queryable report session without replaying the capture. The directory must contain FRAME.xls and GPUTRACE_FRAME.xls, directly or in a child auto-export folder.

List queryable GPU Trace reports currently registered in this MCP process.

Page through every exported GPU Trace metric for the full frame or a marker range. Metrics preserve their raw Nsight names and numeric values and add inferred display groups and units.

Page through exported marker/regime ranges with hierarchy, path, parent id, and measured duration. A returned range_id can be used by the metric, summary, and shader-occupancy tools.

Return the same Top-Level Throughput, Active Threads Per Warp, and SM Warp Occupancy summary for the full trace or one exported marker range.

Read the exported Trace Information name/value table, including collection and replay configuration. Credential-shaped values are redacted.

Page through per-shader PC-sampler occupancy rows for the full trace or one marker range. Returns the exact exported shader name/hash and aggregate warps/occupancy; stage is explicitly heuristic.

Inspect the retained .ngfx-gputrace WRPV container and return frame timing, queues, command-stream/API-call counts, timestamp coverage, shader-profiler inventory, and the version-matched embedded protobuf schema inventory.

Page through queue command-stream calls decoded from .ngfx-gputrace, including queue identity and timestamp intervals. Coalesced intervals explicitly identify calls that share one measured timing group.

Return one decoded GPU Trace command-stream call with full typed arguments, return value, object handles, timestamp boundaries, pipeline stages, and HES correlation ids. Embedded byte arrays are represented by byte length, not dumped.

Page through D3D12 BeginEvent/EndEvent ranges decoded from the binary report, including full marker path, hierarchy, absolute frame-relative start/end time, timing uncertainty bounds, and the paired timeline event ids. These marker ids can directly scope PC-sampling analysis.

Page through Shader Pipelines data decoded from the binary report and joined with raw PC samples and exported metrics: pipeline type/name, shader stage/hash, exact sample count/share, DXIL or SPIR-V correlation, code address/size, registers, group size, barriers, shared memory, and debug-info availability. An explicitly named estimate is also returned as fallback/comparison.

Return the Shader Pipelines table grouped by Shader Object: exact shader stage/name/hash, raw PC-sample count and Samples %, DXIL/SPIR-V correlation, debug-blob reference state, and referencing pipelines. A blob reference does not imply complete source/call-stack debug information. Sample counts cover Entire Trace.

Decode raw PC-sampling records into sample/stall reasons, exact shader/code-block attribution, and top shader and instruction-PC hotspots. Scope the query to the entire trace, a frame-relative time window, or a timed marker id such as BasePass. This does not require shader PDBs; source-line and SASS text correlation are not claimed.

Unregister a GPU Trace report. Temporary artifacts created without output_directory are deleted; user-selected persistent output directories are never deleted.

Replay the capture in a hidden, vsync-off window and summarize Nsight's per-iteration CPU submit, finish-wait, reset, FPS, screenshot-difference, and optional GPU frame-time measurements. This executes GPU work; it is not a static read.

Clear in-memory sanitized Nsight export data held by this MCP process.