AI coding-agent skills (plus a subagent) for zymtrace, the continuous CPU and GPU profiling and optimization platform. Profile and optimize the efficiency of your CPU and GPU-accelerated workloads (find the bottleneck and apply the fix), and install, upgrade, expose, and troubleshoot the platform. Works in Claude Code, OpenAI Codex, and Cursor from the same source.
📚 Docs: https://docs.zymtrace.com/ai-agent-skills/
Once installed, describe what you want in plain English and your coding agent handles the rest:
"install zymtrace on my eks cluster"
"upgrade zymtrace to 26.5.0"
"expose the gateway via internal alb"
"install the profiler with gpu support"
"no profiles showing up in the ui. why?"
"connect zymtrace mcp so I can analyze my profiles from here"
"analyze my vllm workload, find the bottleneck"
"analyze the gpu flamegraph over the last 1 hour and suggest solutions" # ← the universal starter
zymtrace-agent-skills.webm
| Skill | What it does |
|---|---|
install-zymtrace-backend |
Install the backend on Kubernetes (Helm) or Docker Compose. Handles license, databases, air-gapped registries. |
upgrade-zymtrace-backend |
Bump image tags, chart versions, or both. Built-in rollback and verification. |
expose-zymtrace-backend |
Expose the gateway via NodePort, AWS ALB, NGINX Ingress, or cloud LoadBalancer. TLS included. |
install-zymtrace-profiler |
Install the agent on Kubernetes, Docker, or bare-metal. CPU + CUDA GPU profiling (CUDA 12.x+). |
troubleshoot-zymtrace-backend |
Diagnose "no data appearing", license errors, ingest crashes, slow queries, storage issues. |
troubleshoot-zymtrace-profiler |
Diagnose agent-side failures: CrashLoopBackOff, OOMKilled, NVML missing, PC sampling, license rejected. |
configure-zymtrace-mcp |
Connect Claude Code (or any MCP client) to the zymtrace MCP server so you can analyze profiles with natural-language queries. |
optimize-cpu-workloads |
Analyze CPU performance through the MCP (CPU-only deployments): rank the top consumers (which process/function/pod is hottest, biggest ROI) or drill into a named workload, pull CPU metrics + flamegraph, name the pattern, scope to the user's own code, then apply the fix in the local source (asking for the path if it isn't local) and close with a follow-up. No GPU or inference-server framing. |
optimize-gpu-workloads |
Analyze GPU performance through the MCP: read a GPU flamegraph, find the hot kernel, diagnose low GPU saturation, or investigate vLLM/SGLang/Triton or training. Always cross-checks the matching CPU flamegraph (the bottleneck often hides host-side), then applies the fix in the local source and closes with a follow-up. |
optimize-memory-allocation |
Analyze JVM memory allocation profiles through the MCP: what's allocating the most (bytes/objects), why GC is busy, where churn burns CPU. Java-only (boxing, String churn, collection resize, buffer churn). Names the allocation site, applies the fix, frames impact as GC pause + CPU reclaimed. |
| Agent | What it does |
|---|---|
zymtrace-perf-engineer |
Autonomous, hands-off performance investigation. Ranks the top consumers (which process/app is hottest, biggest ROI) or identifies the named entity (script for Python, container, host, or k8s pod/deployment), pulls its metrics, then the CPU flamegraph (and the GPU flamegraph when it's a GPU workload), then applies the fix in the local source (asking for the path if it isn't local) and closes with a follow-up, all without stopping to confirm each step. Runs several in parallel. Invoke it by name (e.g. "use the zymtrace-perf-engineer to analyze my vLLM GPU workload"). |
The same skills install into any of these agents. The repo carries a per-tool manifest; there's one canonical copy of the skills underneath.
| Tool | Install |
|---|---|
| Claude Code | claude plugin marketplace add zystem-io/zymtrace-skills → claude plugin install zymtrace@zymtrace-skills |
| OpenAI Codex | codex plugin marketplace add zystem-io/zymtrace-skills, then install zymtrace via /plugins |
| Cursor | Settings → Plugins → Team Marketplaces → Import → paste the repo URL |
Inside any Claude Code session (or your terminal), run:
claude plugin marketplace add zystem-io/zymtrace-skills
claude plugin install zymtrace@zymtrace-skillsThat's it. Skills become available as /zymtrace:install-zymtrace-backend, /zymtrace:upgrade-zymtrace-backend, and so on. (claude plugin marketplace list confirms the marketplace was added, but it's the install step that enables the skills.)
codex plugin marketplace add zystem-io/zymtrace-skillsThen run /plugins in Codex and install zymtrace from the marketplace.
- Open Dashboard → Settings → Plugins.
- Under Team Marketplaces, click Import.
- Paste the repository URL
https://github.com/zystem-io/zymtrace-skillsand continue. - Review the parsed zymtrace plugin, set access/name as you like, and save.
Describe what you want, and Claude Code routes to the right skill automatically. Or invoke a skill directly with its name:
/zymtrace:install-zymtrace-backend
/zymtrace:upgrade-zymtrace-backend
/zymtrace:expose-zymtrace-backend
/zymtrace:install-zymtrace-profiler
/zymtrace:troubleshoot-zymtrace-backend
Each skill walks you through the decisions, runs the right commands, and verifies the result. You stay in the driver's seat; every change is confirmed with you first.
For an investigation you want to run unattended, hand the whole thing to the zymtrace-perf-engineer agent. Name it in your request:
"use the zymtrace-perf-engineer to analyze my vLLM GPU workload over the last hour"
"have the zymtrace-perf-engineer profile the inference deployment and report back"
It identifies the entity, pulls metrics, then the flamegraph(s), and returns one finished recap, without stopping to confirm each step. Run several at once to triage multiple workloads in parallel. Run /agents to see it in the live list. Unlike the skills, it runs autonomously, so the read-only profile pulls don't prompt for each step.
optimize-cpu-workloads turns CPU cores into money: it annualizes each workload's cores and shows an estimated yearly cost. By default it assumes $0.04 / vCPU-hour (≈ $350 / core-year) and labels it as an assumption. To use your real number (your blended instance rate, a Savings-Plan effective rate, or a fully-loaded chargeback figure), add one line to your agent-instructions file, either AGENTS.md (Claude Code, Codex, Cursor) or CLAUDE.md (Claude Code):
zymtrace-vcpu-rate: $0.032/vCPU-hourThe skill checks for that line before falling back to the default, so every cost estimate uses your rate and it stops asking. (You can also tell it your rate mid-session and approve saving it; it writes the same line for you.)
Run /skills and /agents in any session to see the components with on/off toggles. For the full inventory and per-component token cost, run:
claude plugin details zymtraceComponent inventory
Skills (10) optimize-cpu-workloads, optimize-gpu-workloads,
optimize-memory-allocation, configure-zymtrace-mcp, expose-zymtrace-backend,
install-zymtrace-backend, install-zymtrace-profiler, troubleshoot-zymtrace-backend,
troubleshoot-zymtrace-profiler, upgrade-zymtrace-backend
Agents (1) zymtrace-perf-engineer
The command also prints a live per-component token-cost breakdown (always-on vs on-invoke).
Clone the repo and install the plugin from the local checkout. It loads as a plugin, so ${CLAUDE_PLUGIN_ROOT} and the helper scripts resolve:
git clone https://github.com/zystem-io/zymtrace-skills.git
cd zymtrace-skills
claude plugin validate ./zymtrace # fast check: manifest parses (plugin root is zymtrace/)
claude plugin marketplace add "$PWD" # register the repo-root marketplace (zymtrace-skills)
claude plugin install zymtrace@zymtrace-skills # install from it; restart to applyAfter editing, re-read the local manifest and restart: claude plugin marketplace update zymtrace-skills.
Or symlink as personal skills (no ${CLAUDE_PLUGIN_ROOT}, so bundled scripts won't resolve):
mkdir -p ~/.claude/skills
ln -s "$PWD"/zymtrace/skills/*/ ~/.claude/skills/Run the tests before committing:
python -m venv .venv && source .venv/bin/activate
make install # pytest + PyYAML
make test # structural tests: layout, frontmatter, version sync, path checksThe structural suite needs no API keys, cluster, or network, and runs in CI on every branch. See CLAUDE.md for repo conventions and how to add a skill.
Once your backend is up and the profiler is reporting, connect the zymtrace MCP and analyze GPU and CPU flamegraphs straight from your agent.
- AI Agent Skills docs: https://docs.zymtrace.com/ai-agent-skills/
- MCP setup: https://docs.zymtrace.com/mcp