# Terradev

> Cross-cloud GPU orchestration CLI.

- name: io.github.theoddden/terradev
- version: 6.2.20
- connection class: R0 (autonomous)
- trust: 90/100
- quality: 57/100 (needs work)
- owner: not claimed
- descriptor: https://api.protogrid.dev/v1/servers/io.github.theoddden%2Fterradev
- tools: https://api.protogrid.dev/v1/servers/io.github.theoddden%2Fterradev/tools

**An agent can connect right now, no human step.** Remote endpoint, no authentication, reachable on the last probe.

## Connection (mcpServers)

```json
{
  "mcpServers": {
    "terradev": {
      "type": "sse",
      "url": "https://terradev-mcp.terradev.cloud/sse"
    }
  }
}
```


## Trust

- hygiene: 100
- liveness: 100
- freshness: 85
- provenance: 75
- drivers: +repository +namespace-matches-repo -thin-description +reachable +uptime-100% ~updated-40d-ago
- Derived from observable signals (official registry feed and our own credential-free probes); no code audit performed.

## Quality 57/100 (needs work)

### protocol: 50 (weight 25)

- warn `protocol.modern`: newest supported version is 2025-11-25; 2026-07-28 not supported, older versions are deprecated until 2027-07-28
- n/a `protocol.stateless`: only applies to 2026-07-28 servers
- warn `protocol.transport`: only the legacy HTTP+SSE transport, deprecated until 2027-07-28
- n/a `protocol.list_ttl`: only applies to 2026-07-28 servers

### authorization: n/a (weight 15)

- n/a `auth.prm`, `auth.as_metadata`, `auth.cimd`: no remote uses OAuth
- n/a `auth.secret_in_url`: no templated URL

### tool hygiene: 50 (weight 30)

- warn `tools.descriptions`: 3 of 277 tools have a description under 20 characters
- pass `tools.description_length`: descriptions are concise
- pass `tools.schemas`: every tool has an object input schema
- fail `tools.annotations`: no tool declares readOnlyHint or destructiveHint, so clients cannot tell safe calls from risky ones
- fail `tools.directory_hints`: no tool declares all four of readOnlyHint, destructiveHint, idempotentHint and openWorldHint; missing: provision_gpu (readOnlyHint, destructiveHint, idempotentHint, openWorldHint); local_scan (readOnlyHint, destructiveHint, idempotentHint, openWorldHint); governance_request_consent (readOnlyHint, destructiveHint, idempotentHint, openWorldHint); …
- fail `tools.token_cost`: about 38,007 tokens to load every tool, a large share of an agent's context
- pass `tools.api_dump`: tools are not a one-to-one API dump

### stability: 75 (weight 15)

- warn `stability.changes`: 32 tool changes in 30 days
- pass `stability.rug_pull`: no tool changed its meaning under the same name

### dependencies: 67 (weight 15)

- fail `deps.known_vulns`: 2 advisories to act on (malicious, or rated high or critical with a fix available): GHSA-2g87-pp6c-29x5 in aiohttp@3.14.3 (direct, fixed in 3.14.4), GHSA-5wr8-gcqf-gqf3 in aiohttp@3.14.3 (direct, fixed in 3.14.4); 1 more of lower severity or without a fix
- pass `deps.mcp_sdk_version`: mcp 1.30.0, no known advisory
- pass `deps.resolvable`: terradev-cli@6.2.20 resolved: 66 packages

- tools: 277, about 38,007 tokens to load them all
- tool set hash: 9f07aeadf9214fe140f44e6cbd58a03c
- tools last changed: 2026-09-29T04:57:51.122Z
- badge: [![protogrid quality](https://protogrid.dev/badge/io.github.theoddden/terradev.svg)](https://protogrid.dev/servers/io.github.theoddden/terradev)
- Checks run on what our credential-free, read-only probes observe; tools are never called and no code audit is performed.

### Tool changes

- 2026-09-29 `triggers_history` added
  - Show trigger execution history.
- 2026-09-29 `triggers_execute` added
  - Manually execute a trigger now.
- 2026-09-29 `triggers_enable` added
  - Enable or disable a trigger.
- 2026-09-29 `triggers_list` added
  - List configured triggers.
- 2026-09-29 `triggers_create` added
  - Create an event trigger (schedule/webhook/metric).
- 2026-09-29 `quote_gpu` added
  - Get a GPU price quote across providers.
- 2026-09-29 `preflight_provision` added
  - Run preflight validation before provisioning GPU instances.
- 2026-09-29 `migrate_compare` added
  - Compare migration target providers by cost/compatibility.
- 2026-09-29 `migrate_rollback` added
  - Roll back a migration to the source provider.
- 2026-09-29 `migrate_status` added
  - Show status of running/completed migrations.

Full history: https://api.protogrid.dev/v1/servers/io.github.theoddden%2Fterradev/changes

## Dependencies

- pypi terradev-cli@6.2.20: resolved, 65 dependencies, MCP SDK mcp 1.30.0
  - GHSA-2g87-pp6c-29x5 (high) in `` aiohttp@3.14.3 ``, direct, fixed in `` 3.14.4 ``, advisory summary: "aiohttp: Multiple instances of excessive or unbounded memory/CPU use" https://osv.dev/vulnerability/GHSA-2g87-pp6c-29x5
  - GHSA-5wr8-gcqf-gqf3 (high) in `` aiohttp@3.14.3 ``, direct, fixed in `` 3.14.4 ``, advisory summary: "aiohttp server lacks an initial request/header timeout" https://osv.dev/vulnerability/GHSA-5wr8-gcqf-gqf3
  - GHSA-494j-54xm-rwp3 (low) in `` aiohttp@3.14.3 ``, direct, fixed in `` 3.14.4 ``, advisory summary: "AIOHTTP: Secure attribute not enforced for shared cookie" https://osv.dev/vulnerability/GHSA-494j-54xm-rwp3

Dependency graphs from deps.dev; advisories from OSV.dev, including the GitHub Advisory Database and the PyPI Advisory Database; all CC BY 4.0.

## Remotes

- https://terradev-mcp.terradev.cloud/sse (sse, auth none, reachable true, uptime30d 1)

## Packages

- pypi terradev-cli@6.2.20 secrets: RUNPOD_API_KEY, VAST_API_KEY, LAMBDA_API_KEY, HF_TOKEN, AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, TERRADEV_MCP_BEARER_TOKEN

## Tools (277)

- `active_context`: Get current Terradev state: running training jobs, active instances, spend-to-date, alerts. Call this on session start to resume context from previous sessions.
- `agent_agentic_serving_configure`: Configure agentic inference serving settings.
- `agent_agentic_serving_helm_values`: Print Helm values for agentic inference deployment.
- `agent_agentic_serving_k8s`: Print K8s deployment manifests for agentic inference.
- `agent_agentic_serving_launch_args`: Print engine launch arguments for copy-paste.
- `agent_agentic_serving_lmcache_env`: Print LMCache environment variables.
- `agent_agentic_serving_show_config`: Show current agentic serving configuration.
- `agent_cost`: Show real-time cost breakdown for a fleet by tier.
- `agent_deploy`: Provision a heterogeneous agent fleet across all tiers simultaneously.
- `agent_langchain_create_langgraph`: Create a LangGraph workflow.
- `agent_langchain_create_pipeline`: Create an SGLang pipeline.
- `agent_langchain_create_workflow`: Create a LangChain workflow.
- `agent_langchain_test`: Test connection to LangChain service.
- `agent_langgraph_create_workflow`: Create a LangGraph workflow.
- `agent_langgraph_deploy`: Deploy a workflow.
- `agent_langgraph_status`: Get workflow status.
- `agent_langgraph_test`: Test connection to LangGraph service.
- `agent_letta_chat`: Send a message to a Letta agent.
- `agent_letta_create`: Create a new stateful Letta agent.
- `agent_letta_delete`: Delete a Letta agent.
- `agent_letta_list`: List Letta agents.
- `agent_letta_remember`: Teach a Letta agent a durable fact.
- `agent_letta_status`: Show the state of a Letta agent.
- `agent_list`: List all known agent fleets.
- `agent_mem0_add`: Store a memory in Mem0 for an agent or user.
- `agent_mem0_configure`: Configure Mem0 credentials and defaults.
- `agent_mem0_delete`: Delete a memory by ID.
- `agent_mem0_forget`: Delete all memories matching an entity scope.
- `agent_mem0_get`: Get a single memory by ID.
- `agent_mem0_list`: List memories for an entity scope.
- `agent_mem0_search`: Search agent/user memories.
- `agent_mem0_test`: Test connection to Mem0.
- `agent_mem0_update`: Update a memory by ID.
- `agent_plan`: Plan a heterogeneous agent fleet without provisioning.
- `agent_scale`: Scale a single fleet tier up or down without affecting other tiers.
- `agent_skill_attach`: Attach a skill.md to a Letta agent as a durable memory block.
- `agent_skill_init`: Create a skill.md template for an agent.
- `agent_status`: Show live status of a fleet — tier health, KV hit rate, queue depth, cost.
- `agent_teardown`: Terminate all fleet instances and remove fleet state.
- `agent_vector_db_down`: Teardown a vector database provisioned for an agent fleet.
- `agent_vector_db_up`: Provision a vector database for an agent fleet.
- `analytics`: Get cost analytics
- `checkpoint_delete`: Delete a checkpoint.
- `checkpoint_list`: List all checkpoints for a training job.
- `checkpoint_promote`: Promote a checkpoint to a final model path for serving.
- `checkpoint_restore`: Restore a specific checkpoint for a training job.
- `checkpoint_save`: Manually trigger a checkpoint save for a running training job.
- `configure_provider`: Configure provider credentials
- `cost_analyze`: Deep cost analysis of current GPU infrastructure: per-provider breakdown, utilization efficiency, waste identification, and optimization potential.
- `cost_optimize_recommend`: Generate actionable cost optimization recommendations: spot migration, GPU right-sizing, provider arbitrage, idle shutdown, and density packing.
- `cost_simulate`: Simulate cost optimization scenarios with ROI projections. Compare current vs optimized infrastructure costs.
- `create_postgresql_connection`: Create a PostgreSQL database connection with auto-table creation. Returns a connection ID for subsequent operations.
- `create_sqlite_connection`: Create a SQLite database connection with auto-table creation. Returns a connection ID for subsequent operations.
- `database_weaviate_create_collection`: Create a Weaviate collection.
- `database_weaviate_delete_collection`: Delete a Weaviate collection.
- `database_weaviate_hybrid_search`: Hybrid vector + BM25 search in a Weaviate collection.
- `database_weaviate_insert`: Insert objects into a Weaviate collection.
- `database_weaviate_list_collections`: List Weaviate collections.
- `database_weaviate_query`: Vector similarity search in a Weaviate collection.
- `database_weaviate_up`: Initialize a Weaviate connection.
- `deepeval_evaluate`: Evaluate a single LLM output with a DeepEval metric (AnswerRelevancy, Faithfulness, Hallucination, etc.).
- `deepeval_init`: Generate a starter DeepEval test file.
- `deepeval_metrics`: List available DeepEval metrics for LLM evaluation.
- `deepeval_run`: Run a DeepEval test suite from a Python test file.
- `dvc_diff`: Show DVC diff between two revisions (e.g. training checkpoints). Shows added, modified, deleted files.
- `dvc_push`: Push DVC-tracked data to the configured remote storage.
- `dvc_stage_checkpoint`: Atomic checkpoint staging: DVC add + push + git commit in one operation. Promotes a training checkpoint to versioned storage.
- `dvc_status`: Get DVC repository status: tracked files, remotes, and changes since last commit.
- `egress_cheapest_route`: Find the cheapest egress route between cloud providers/regions for model weights or dataset transfer. Supports multi-hop routing.
- `egress_optimize_staging`: Optimize dataset or model staging across regions by finding the cheapest transfer plan. Integrates with the dataset stager for parallel uploads.
- `environments_approve`: Approve a pending environment promotion.
- `environments_create`: Create a named deployment environment (dev/staging/prod) with config.
- `environments_history`: Show promotion/change history for an environment.
- `environments_list`: List deployment environments.
- `environments_promote`: Promote a workload or config between environments.
- `get_database_connection`: Get information about a database connection including type, status, and configuration.
- `governance_compliance_report`: Generate comprehensive compliance report: consent stats, policy evaluations, data movements, violations. For GDPR/SOC2/HIPAA audits.
- `governance_evaluate_opa`: Evaluate OPA (Open Policy Agent) policies for data access. Checks region restrictions, classification rules, and compliance requirements.
- `governance_move_data`: Move data with full governance audit trail. Requires prior consent and OPA policy approval. Tracks integrity, encryption, and compliance.
- `governance_movement_history`: Get data movement audit log. Filter by user, dataset, or time range.
- `governance_record_consent`: Record a consent response (granted or denied) for a pending consent request.
- `governance_request_consent`: Request user consent for data movement across cloud regions. GDPR/SOC2 compliant consent tracking with audit trail.
- `gpu_topology`: GPU NUMA topology report with intra-GPU XCD (Accelerated Compute Die) awareness. Models MI300X (8 XCDs, 192GB HBM3), MI300A (6 XCDs, 128GB), H200 (unified 141GB HBM3e), H100 (80GB). Reports PCIe locality (PIX/PXB/PHB/SYS), GPU-NIC pairing, SR-IOV VF status, and generates XCD-aware NCCL/AITER environment variables.
- `guardrails_chat`: Send a message through NeMo Guardrails and return the safety-filtered response. Applies topical, jailbreak, PII, and factcheck rails.
- `guardrails_generate_config`: Generate default Colang 2.x guardrails configuration files (topical, jailbreak, PII, factcheck rails).
- `guardrails_k8s`: Generate Kubernetes deployment manifest for NeMo Guardrails server (standalone or sidecar mode).
- `guardrails_test`: Test connection to NeMo Guardrails server.
- `helm_generate`: Generate Helm charts from workload specifications.
- `hf_create_endpoint`: Create a HuggingFace Inference Endpoint (paid GPU endpoint). Supports custom GPU types, regions, and scaling.
- `hf_delete_endpoint`: Delete a HuggingFace Inference Endpoint.
- `hf_endpoint_infer`: Run inference on a HuggingFace Inference Endpoint. Supports text generation, embeddings, and custom inputs.
- `hf_endpoint_info`: Get detailed info about a specific HuggingFace Inference Endpoint: status, URL, scaling config, cost.
- `hf_hardware_compare`: Compare all hardware options for a HuggingFace model. Returns side-by-side cost, performance, and compatibility analysis.
- `hf_hardware_recommend`: Get hardware recommendation with cost breakdown for any HuggingFace model. Returns optimal GPU type, estimated cost, and performance score.
- `hf_list_datasets`: Search and browse HuggingFace Hub datasets. Filter by author and search query.
- `hf_list_endpoints`: List all active HuggingFace Inference Endpoints with status, URL, and cost.
- `hf_list_models`: Search and browse HuggingFace Hub models. Filter by author, task, library. Returns model ID, downloads, likes, and tags.
- `hf_model_info`: Get detailed model info: architecture, size, downloads, license, tags, pipeline_tag, and model card.
- `hf_smart_template`: Auto-generate an optimized deployment template for any HuggingFace model. Analyzes model size, architecture, and quantization to select optimal hardware and generate ready-to-deploy configs.
- `hf_space_deploy`: Deploy model to HuggingFace Spaces
- `hf_space_status`: Get HuggingFace Space deployment status.
- `infer_failover`: Run health checks and auto-failover for inference endpoints. If a primary endpoint is unhealthy and has a backup configured, traffic automatically shifts to the backup provider.
- `infer_route`: Semantic-aware inference routing. Analyzes query content across 6 signal dimensions (modality, complexity, domain, language, safety, keywords), applies NUMA-aware endpoint scoring, and selects the optimal inference endpoint. Uses DAG parallel execution for signal extraction.
- `infer_route_disagg`: Disaggregated Prefill/Decode routing (DistServe architecture). Splits LLM inference into compute-bound prefill phase (routed to FLOPS-optimized GPUs like H100 SXM) and memory-bound decode phase (routed to bandwidth-optimized GPUs like MI300X). Tracks KV cache handoffs between endpoint pairs.
- `inferx_configure`: Configure InferX serverless platform credentials.
- `inferx_delete`: Delete an InferX model deployment.
- `inferx_deploy`: Deploy model to InferX serverless platform
- `inferx_list`: List deployed InferX models
- `inferx_optimize`: Get cost analysis for inference endpoints
- `inferx_quote`: Get InferX pricing quotes for a GPU type.
- `inferx_status`: Check InferX endpoint status
- `inferx_usage`: Get InferX account usage statistics: requests, cost, GPU hours, latency.
- `k8s_create`: Create Kubernetes cluster with GPU nodes for optimal multi-cloud deployment
- `k8s_destroy`: Destroy a Kubernetes cluster
- `k8s_device_plugin`: Configure Kubernetes GPU device plugin settings: time-slicing, MIG strategy, and resource naming.
- `k8s_gpu_operator_install`: Install NVIDIA GPU Operator on a Kubernetes cluster. Configures driver containers, device plugin, DCGM exporter, and GPU Feature Discovery.
- `k8s_info`: Get information about a specific cluster
- `k8s_list`: List Kubernetes clusters
- `k8s_mig_configure`: Configure Multi-Instance GPU (MIG) partitioning on A100/H100 GPUs. Splits a single GPU into isolated instances for multi-tenant workloads.
- `k8s_time_slicing`: Configure GPU time-slicing for Kubernetes. Allows multiple pods to share a single GPU with configurable oversubscription.
- `karpenter_create_nodepool`: Create a Karpenter GPU NodePool.
- `karpenter_delete_nodepool`: Delete a Karpenter NodePool.
- `karpenter_events`: Show recent Karpenter scaling events.
- `karpenter_gpu_nodes`: List GPU nodes provisioned by Karpenter.
- `karpenter_install`: Install Karpenter on an EKS cluster for GPU node autoscaling.
- `karpenter_logs`: Show Karpenter controller logs.
- `karpenter_nodepools`: List Karpenter NodePools.
- `karpenter_resources`: Show Karpenter resource usage and capacity.
- `karpenter_status`: Show Karpenter installation/controller status.
- `kserve_generate_yaml`: Generate a GPU-aware KServe InferenceService YAML manifest with NUMA pinning, resource limits derived from model size and VRAM, and topology hints.
- `kserve_list`: List KServe InferenceServices in a Kubernetes namespace.
- `kserve_status`: Get detailed status of a KServe InferenceService including readiness, traffic split, and URL.
- `kv_cache_efficiency`: Analyze KV cache efficiency across serving endpoints.
- `langchain_create_sglang_pipeline`: Create an SGLang model-serving pipeline via LangChain. Connects LangChain agents to SGLang inference endpoints.
- `langchain_create_workflow`: Create a LangChain workflow.
- `langfuse_configure`: Configure Langfuse credentials (public key, secret key, host URL).
- `langfuse_datasets`: List Langfuse datasets for evaluation and fine-tuning.
- `langfuse_export_training_data`: Export Langfuse traces as instruction/response pairs for LoRA fine-tuning. Filters by quality score.
- `langfuse_k8s`: Generate Kubernetes deployment manifest for self-hosted Langfuse.
- `langfuse_otel_env`: Print OTEL environment variables for instrumenting LLM apps to send traces to Langfuse.
- `langfuse_quality`: Get aggregated quality metrics from Langfuse scores for drift detection.
- `langfuse_score`: Create an evaluation score for a Langfuse trace (e.g. quality, accuracy, relevance).
- `langfuse_scores`: List evaluation scores from Langfuse, optionally filtered by trace or score name.
- `langfuse_test`: Test Langfuse connectivity and list accessible projects.
- `langfuse_trace`: Get a single Langfuse trace with all observations/spans.
- `langfuse_traces`: List recent LLM traces from Langfuse.
- `langgraph_create_workflow`: Create a LangGraph stateful workflow with monitoring. Supports agent graphs, tool calling, and state persistence.
- `langgraph_evaluation_workflow`: Create an evaluator-optimizer workflow in LangGraph. Generates outputs, evaluates quality, and iteratively improves.
- `langgraph_orchestrator_worker`: Create an orchestrator-worker pattern workflow in LangGraph. The orchestrator delegates tasks to specialized worker agents.
- `langgraph_workflow_status`: Get the status and metrics of a LangGraph workflow execution.
- `lineage_checkpoint`: Record a data/model lineage checkpoint.
- `lineage_diff`: Diff lineage between two artifacts or checkpoints.
- `lineage_export`: Export lineage data (JSON/graph) for an artifact.
- `lineage_graph`: Render the lineage graph for an artifact.
- `lineage_trace`: Trace the lineage of a dataset or model artifact.
- `local_scan`: Scan local machine and network for available GPU devices. Returns total VRAM pool for local-first provisioning.
- `lora_add`: Hot-load a LoRA adapter onto a running vLLM endpoint. The adapter becomes immediately available as a model name for inference requests. Uses vLLM's fused_moe_lora kernel for 454% higher output tokens/sec on MoE models.
- `lora_list`: List LoRA adapters loaded on a running vLLM endpoint. Shows base models and hot-loaded fine-tuned adapters.
- `lora_remove`: Hot-unload a LoRA adapter from a running vLLM endpoint. Frees GPU memory for other adapters.
- `manage_instance`: Manage GPU instances (stop/start/terminate)
- `manifests`: List cached manifests and versions for jobs.
- `migrate_compare`: Compare migration target providers by cost/compatibility.
- `migrate_execute`: Execute a previously-planned workload migration.
- `migrate_plan`: Plan a workload migration between providers with cost analysis.
- `migrate_rollback`: Roll back a migration to the source provider.
- `migrate_status`: Show status of running/completed migrations.
- `mlflow_list_experiments`: List MLflow experiments on the configured tracking server.
- `mlflow_log_run`: Log a Terradev training run to MLflow with auto-injected GPU type, provider, cost/hr, and duration as params.
- `mlflow_register_model`: Register a trained model in the MLflow model registry with Terradev provenance tags.
- `ml_vllm_lora_link`: Load the active registry version of an adapter onto a vLLM server.
- `ml_vllm_lora_list`: List LoRA adapters currently loaded on a vLLM server.
- `ml_vllm_lora_load`: Hot-load a LoRA adapter onto a running vLLM server.
- `ml_vllm_lora_sync`: Synchronize an adapter from the registry across multiple vLLM replicas.
- `ml_vllm_lora_unload`: Hot-unload a LoRA adapter from a running vLLM server.
- `moe_deploy`: Deploy Mixture-of-Experts models with production-ready cluster templates. Auto-applies vLLM cost optimizations (KV cache offloading for up to 9x throughput, MTP speculative decoding for up to 2.8x speed, sleep mode for 18-200x faster restarts). Supports GLM-5, Qwen 3.5, Mistral Large 3, DeepSeek V4, Llama 5. Configures NVLink topology, tensor parallelism, FP8 quantization, vLLM/SGLang backends, and GPU-aware HPA autoscaling.
- `ollama_chat`: Chat with an Ollama model using the chat/completions API.
- `ollama_generate`: Generate text using an Ollama model (non-chat completions).
- `ollama_list`: List models available on an Ollama server.
- `ollama_model_info`: Get detailed information about an Ollama model (parameters, template, license).
- `ollama_ps`: List currently running Ollama models.
- `ollama_pull`: Pull a model to an Ollama server on a remote instance.
- `optimize`: Find cheaper alternatives for running instances
- `orchestrator_evict`: Evict a model from GPU memory.
- `orchestrator_infer`: Test inference with a model via the orchestrator.
- `orchestrator_load`: Load a model into GPU memory.
- `orchestrator_register`: Register a model with the orchestrator.
- `orchestrator_start`: Start the model orchestrator for multi-model GPU sharing with eviction policies.
- `orchestrator_status`: Get orchestrator and model status including GPU memory utilization.
- `phoenix_k8s`: Generate Kubernetes deployment manifest for self-hosted Arize Phoenix server.
- `phoenix_otel_env`: Generate OpenTelemetry environment variables for instrumenting serving pods with Phoenix tracing.
- `phoenix_projects`: List Phoenix projects (trace namespaces).
- `phoenix_snippet`: Generate Python instrumentation snippet for adding Phoenix tracing to LLM applications.
- `phoenix_spans`: List recent spans for a Phoenix project. Supports SpanQuery DSL filters like "span_kind == 'RETRIEVER'" or "status_code == 'ERROR'".
- `phoenix_test`: Test connection to Arize Phoenix server. Returns collector endpoint and project count.
- `phoenix_trace`: View full execution tree for a specific trace ID. Shows span hierarchy, latencies, and token counts.
- `preflight`: Pre-training validation: GPU availability, NCCL, RDMA, drivers across all nodes.
- `preflight_gpu_check`: GPU-specific preflight validation: NVIDIA drivers, CUDA version, GPU count, NCCL, NVLink topology, NCU stall-signature profiling, and adversarial config verification (V1-V3).
- `preflight_network_check`: Network-specific preflight validation: RDMA availability, InfiniBand status, inter-node bandwidth, latency matrix, firewall rules.
- `preflight_provision`: Run preflight validation before provisioning GPU instances.
- `preflight_report`: Generate full preflight validation report with pass/warn/fail per check. Covers GPU drivers, CUDA, NCCL, RDMA, network, disk, and Docker.
- `price_intel`: GPU price intelligence with quantitative analytics. Computes delta (rate of change), gamma (acceleration), and annualized realized volatility on GPU spot/on-demand prices across 21+ providers. Identifies cheapest time windows and provider arbitrage opportunities.
- `price_spot_risk`: Spot instance risk assessment per provider. Returns interruption probability, mean time to interruption, and recommended mitigation.
- `price_trends`: Get GPU price trend analysis with delta (rate of change), gamma (acceleration), and annualized volatility. Identifies cheapest time windows.
- `provision_gpu`: Provision GPU instances for optimal parallel efficiency
- `qdrant_collections`: List all Qdrant vector collections with their point counts and configurations.
- `qdrant_count`: Count points (vectors) in a Qdrant collection.
- `qdrant_create_collection`: Create a Qdrant vector collection. Auto-configures vector dimensions from embedding model name.
- `qdrant_info`: Get detailed info and stats for a Qdrant collection.
- `qdrant_k8s`: Generate Kubernetes StatefulSet manifest for self-hosted Qdrant vector database.
- `qdrant_test`: Test connection to Qdrant vector database. Returns cluster info and collection count.
- `query_database`: Execute a SELECT query on a database connection. Returns query results as a list of dictionaries.
- `quote_gpu`: Get a GPU price quote across providers.
- `ray_disagg_pd_deploy`: Generate a Ray Serve LLM disaggregated Prefill/Decode deployment. Splits inference into compute-bound prefill and memory-bound decode phases with KV cache transfer via NIXL.
- `ray_list_jobs`: List all running Ray jobs and tasks.
- `ray_parallelism_strategy`: Compute optimal TP/DP/EP parallelism strategy for a given MoE model and GPU count. Returns recommended configuration with rationale.
- `ray_start`: Start a Ray cluster (head node or worker). For distributed ML training and inference.
- `ray_status`: Get Ray cluster status including node count, resources, memory, and running jobs.
- `ray_stop`: Stop the Ray cluster on the current node.
- `ray_submit_job`: Submit a job script to the Ray cluster for distributed execution.
- `ray_wide_ep_deploy`: Generate a Ray Serve LLM Wide-EP (Expert Parallel) deployment for MoE models. Returns Python script and config for distributed MoE serving with EPLB and DeepEP.
- `rollback`: Explicit versioned rollback. Format: job@version (e.g., llama3@v3).
- `run_workflow`: Run a declarative YAML workflow that chains multiple Terradev commands (provision → preflight → train → monitor → checkpoint). Returns step-by-step execution status with cost estimates and confirmation gates for expensive operations.
- `secrets_env`: Print environment-style export lines for a provider. By default values are masked.
- `secrets_get`: Retrieve a stored secret. By default the value is masked.
- `secrets_list`: List stored provider and key names. Values are never shown.
- `secrets_remove`: Remove a provider or a single key from the secret store.
- `secrets_run`: Run a shell command with secrets injected into the environment.
- `secrets_set`: Store a secret in the Terradev secrets store. Values are encrypted at rest.
- `secrets_sync`: Import TERRADEV_* environment variables into the secret store for supported cloud providers.
- `secrets_verify`: Check which providers are fully configured and which keys are missing.
- `setup_provider`: Get setup instructions for a provider
- `sglang`: Complete SGLang optimization stack with workload-specific auto-tuning for 7 workload types: agentic chat, batch inference, low latency, MoE models, PD disaggregated, structured output, and RAG.
- `sglang_inference`: [DEPRECATED] Use sglang action='inference' instead. Test inference against a running SGLang endpoint.
- `sglang_start`: [DEPRECATED] Use sglang action='start' instead. Start an SGLang inference server on a remote instance.
- `sglang_stop`: [DEPRECATED] Use sglang action='stop' instead. Stop an SGLang server on a remote instance.
- `stage`: Compress, chunk, checksum, and position datasets near compute. Supports local paths, S3/GCS URIs, HTTP URLs, and HuggingFace dataset names. Returns staging plan with agent recommendations.
- `status`: View all instances and costs with state optimization
- `train`: Launch distributed training on provisioned GPU nodes. Supports torchrun, deepspeed, accelerate, and megatron. Use from_provision='latest' to auto-resolve node IPs from your last provision command.
- `train_detect_stragglers`: Detect straggler nodes in distributed training. Identifies GPUs with significantly lower utilization that slow the whole job.
- `training_config_generate`: Generate a complete training configuration from a declarative spec. Auto-detects framework, sets optimal parallelism, and configures distributed training.
- `training_launch_distributed`: Full distributed training launch with framework auto-detection, topology validation, and monitoring. Combines preflight + train + monitor in one operation.
- `train_monitor`: Real-time GPU monitoring for training jobs. Shows utilization, memory, temperature, power, and cost.
- `train_resume`: Resume a training job from its latest checkpoint. Rebuilds config with topology revalidation.
- `train_snapshot`: Get complete training monitoring snapshot: GPU metrics (utilization, memory, temp, power), training metrics (loss, grad_norm, lr, throughput), straggler detection, and cost estimate.
- `train_status`: List all training jobs and their state (created, running, completed, failed).
- `train_stop`: Stop a running training job. Kills training processes on all nodes in parallel.
- `train_unsloth_run`: Run an Unsloth local model server.
- `train_unsloth_start`: Start a coding agent backed by Unsloth's local model server.
- `train_unsloth_stop`: Stop a running Unsloth server started with `unsloth run`.
- `triggers_create`: Create an event trigger (schedule/webhook/metric).
- `triggers_enable`: Enable or disable a trigger.
- `triggers_execute`: Manually execute a trigger now.
- `triggers_history`: Show trigger execution history.
- `triggers_list`: List configured triggers.
- `up`: CLI-native provisioning with manifest cache and drift detection. Use --fix-drift to detect and auto-fix drifted infrastructure.
- `upsert_database`: Insert or update data in a database table. Performs upsert operation (insert or update on conflict).
- `vllm_analyze`: Analyze a running vLLM server's workload and return specific optimization recommendations with before/after comparisons.
- `vllm_auto_optimize`: Automatically optimize vLLM configuration by analyzing workload patterns. Selects optimal settings for the 6 critical knobs based on live endpoint metrics or sample request files.
- `vllm_benchmark`: Benchmark a vLLM endpoint with concurrent requests. Returns throughput (req/s), success rate, and total latency.
- `vllm_inference`: Test inference against a running vLLM endpoint (completions or chat).
- `vllm_info`: Get vLLM server info: loaded models, config, and health status.
- `vllm_sleep`: Put a vLLM server to sleep. Level 1: offload to CPU (fast wake). Level 2: discard weights (minimal RAM).
- `vllm_start`: Start a vLLM inference server on a remote instance via SSH/systemd. Supports Multi-LoRA, Sleep Mode, KV Offloading, Speculative Decoding.
- `vllm_stop`: Stop a vLLM server on a remote instance.
- `vllm_wake`: Wake a sleeping vLLM server. For Level 2 sleep, also reloads weights and resets prefix cache.
- `wandb_create_dashboard`: Create a custom W&B dashboard with GPU metrics, training loss, and cost panels.
- `wandb_create_report`: Create a W&B report with custom sections, charts, and narrative text.
- `wandb_create_terradev_alerts`: Auto-create standard Terradev alerts: GPU cost > budget, utilization < 50%, training loss spike, straggler detection.
- `wandb_create_terradev_dashboard`: Auto-create a Terradev-specific W&B dashboard with GPU utilization, cost tracking, training metrics, and infrastructure panels.
- `wandb_create_terradev_report`: Auto-generate a Terradev infrastructure report: GPU costs, provider comparison, training efficiency, and recommendations.
- `wandb_dashboard_status`: Get comprehensive W&B monitoring overview: dashboards, reports, alerts, active runs.
- `wandb_list_projects`: List all Weights & Biases projects for the configured entity.
- `wandb_list_runs`: List runs in a W&B project with status, metrics summary, and config.
- `wandb_run_details`: Get detailed info, metrics, and artifacts for a specific W&B run.
- `wandb_setup_alerts`: Set up custom W&B alerts for GPU metrics: cost thresholds, utilization drops, training anomalies.
- `warm_pool_start`: Start the warm pool manager for intelligent model pre-warming. 5 strategies: traffic_based, time_based, priority_based, cost_optimized, latency_optimized.
- `warm_pool_status`: Get warm pool status: hit rate, cold starts, memory saved, cost saved.