Inference AIops

Audit availability: Real writes require a committed pending audit row and stop before contacting the platform if storage is unavailable. Reads and previews may continue with a warning; new audit rows store outcome summaries rather than full responses.

Disclaimer: Community-maintained open-source project. Not affiliated with, endorsed by, or sponsored by the vLLM or Ray projects or any inference-serving vendor. Product and trademark names belong to their owners. MIT licensed.

Governed AI-ops for GPU inference clusters — vLLM (OpenAI API + Prometheus /metrics) and Ray Serve / Ray Jobs (Ray dashboard), plus the single-process serving engines SGLang and TGI (Text Generation Inference) — with a built-in governance harness: unified audit log, policy engine, token/runaway budget guard, undo-token recording, and descriptive risk-tier labels on every audit row. It parses each engine's Prometheus /metrics directly (no Prometheus server required) and probes the Ray dashboard independently. A bearer token is optional (many stacks run open).

Serving engines. vLLM is the flagship (full Ray Serve control plane: scale, drain, autoscale, LoRA, hot-swap). SGLang and TGI are supported for engine-agnostic observability — health, running-model identity, request-latency metrics, queue depth, and latency RCA — read from each engine's own endpoints and metric names. Being single-process servers, they have no Ray-shaped scale/drain API: those writes return a teaching error pointing you at a real horizontal-scale layer (Ray Serve / Kubernetes / a load balancer).

What it does

The flagship value is root-cause analysis, wrapped in guarded reads and writes:

• diagnose_latency_spike (flagship RCA) — when TTFT/TPOT/e2e latency climbs, it correlates queue depth (running vs waiting), KV-cache pressure / preemptions, and prefix-cache locality into a ranked cause plus the specific knob to turn (add replicas, raise max-num-seqs, fix routing, enlarge KV cache). Every flag is a number, not a black-box verdict.
• diagnose_low_utilization — the inverse: idle GPUs, over-provisioned replicas, or routing that strands a cache-warm replica → what to scale down.
• Prometheus-native — reads vLLM's /metrics endpoint directly; no Prometheus/Grafana deployment needed.
• Governance-grade — the first governance-grade entrant in this niche: audit + budget + risk-tier approval + undo-token + prompt-injection sanitize, with dry-run + double-confirm on the fragile prod ops (scale-down, scale-to-zero, drain, redeploy, hot-swap) the community reports as dangerous.
• Laptop self-test — ~80% of the tool self-tests free: vLLM on a single GPU or CPU-mock + Ray in one local container (ray start --head).

What this tool does, and does not, decide

It delivers inference-cluster operations — reads and writes — accurately and efficiently, and records every one of them. It does not decide whether a write is allowed to happen. That is the agent's judgement, or the permission of the environment you connect it with: restrict the network path so it can only reach the read/metrics endpoints, or run the Ray dashboard without its job-submission API, and the writes fail at the server — the place that actually owns the permission.

So there is no read-only switch, no policy file, no approval gate to configure. The one thing the tool guarantees is that nothing is silent: every call, over MCP and over the CLI alike, lands an audit row in ~/.inference-aiops/audit.db, and destructive writes still capture their before-state and record an inverse where one exists.

Each tool declares a risk_level, kept in agreement with its [READ]/[WRITE] documentation tag by a test, and carried into the audit row as a descriptive tier — so a reviewer can see at a glance that a row was a high-risk scale-to-zero. It is a label, not a gate.

Running a smaller / local model? See agent-guardrails.md — it lists the guardrails this tool enforces for you (so you don't spend prompt budget restating them) and gives a ready-made system prompt for what's left.

Capability matrix (39 MCP tools)

The engine-agnostic group works against any supported engine (including vLLM); use it for SGLang/TGI targets or a uniform view across a mixed fleet. The Ray Serve / cluster / deploy write groups are vLLM-only (Ray control plane) — they teach-and-refuse on a SGLang/TGI target.

23 read, 16 write. High-risk writes (scale_replicas_down, scale_to_zero, drain_replica, lora_unload, model_sleep, replica_restart, model_undeploy, deployment_redeploy) all support dry_run + double-confirm; reversible writes record an undo descriptor.

Sleep Mode requires a dev-mode server. vLLM registers /sleep, /wake_up and /is_sleeping only when started with VLLM_SERVER_DEV_MODE=1. Against any other server these three tools report that the route is absent and why, rather than failing vaguely. Sleep Mode suspends the same model; it does not swap base models — serving a different base model means restarting vLLM with a different --model.

Install

uv tool install inference-aiops          # or: pipx install inference-aiops

Quick start

As a Claude Code plugin

One install gives an agent both the skill and the MCP server:

/plugin marketplace add AIops-tools/marketplace
/plugin install inference-aiops@aiops-tools

The MCP server is fetched with uv and pinned to the package version this plugin declares, so an audit row can be traced back to the code that wrote it. Credentials are still configured with inference-aiops init — see below.

As an OpenClaw plugin

The same bundle is published on ClawHub, where one install delivers the skill and its MCP server together:

openclaw plugins install clawhub:@zw008/inference-aiops
openclaw skills info inference-aiops          # expect: Visible to model: yes

Restart the OpenClaw gateway afterwards so it loads the plugin. The MCP server is fetched with uv, pinned to this exact release, so uvx has to be on PATH — without it the skill still installs but reports Visible to model: no. Credentials are configured exactly as below.

As a CLI or standalone MCP server

inference-aiops init                     # wizard: engine (vllm/sglang/tgi) + host + port + scheme
inference-aiops doctor                   # vLLM: probes Ray + vLLM; SGLang/TGI: engine health + inventory
inference-aiops overview                 # deployments + total replicas + queue backpressure
inference-aiops metrics diagnose         # why is inference slow? ranked RCA + the knob to turn
inference-aiops serve list               # Ray Serve deployments + replica counts

Run as an MCP server (stdio) for the full 39-tool surface:

export INFERENCE_AIOPS_MASTER_PASSWORD=...   # only if a bearer token is stored
inference-aiops mcp

Where that password then lives: an exported variable is readable by every process this shell starts and is recorded by shell history. On a shared or long-lived host, prefer the interactive prompt, or inject it from a secret manager for the life of the one command that needs it.

The CLI is a convenience subset (init, overview, serve …, metrics …, secret …, doctor, mcp); the full 39 tools are exposed via the MCP server.

Governance

Every MCP tool passes through the bundled @governed_tool harness. It does not decide whether a write is permitted — see What this tool does, and does not, decide above — but it records every call:

• Audit — every call (params, outcome summary, status, duration, risk tier, and any approver/rationale annotation) logged to ~/.inference-aiops/audit.db (relocatable via INFERENCE_AIOPS_HOME).
• Budget / runaway guard — a safety backstop, not authorization: token and call budgets trip a circuit breaker on tight poll/retry loops.
• Risk tier — each audit row carries a descriptive tier derived from the tool's risk_level; it is a label, not a gate. INFERENCE_AUDIT_APPROVED_BY / INFERENCE_AUDIT_RATIONALE are optional annotations recorded when set, never required.
• Undo recording — reversible writes (scale, autoscale-config, routing, hot-swap, LoRA load) record an inverse descriptor.

Supported scope + limitations

Behaviour is exercised by the test suite against mocked vLLM /metrics, vLLM OpenAI API, and Ray dashboard responses. ~80% of the tool self-tests on a laptop — vLLM on a single GPU or CPU-mock plus a local one-node Ray head. It has not been run against a live production cluster; see docs/VERIFICATION.md for the live-verification checklist.

Unverified against real hardware / topology:

• multi-GPU tensor-parallel / pipeline-parallel deployments,
• real GPU thermal / throttle telemetry (utilisation is best-effort from the Ray dashboard's /api/nodes),
• multi-node drain and node-reboot orchestration.

The fastest live check is inference-aiops doctor; the full checklist lives in docs/VERIFICATION.md.

Missing a capability?

This is the GPU-inference member of the AIops-tools family (governed AI-ops with audit + budget + undo + risk tiers). If a vLLM or Ray capability you need is missing, or your stack speaks a dialect these tools don't yet handle — open an issue or a PR. Contributions welcome.