Stack chooser
Your AI stack, not the vendor's.
Fifteen layers in six blocks. You need the ones your problems light up.
Six facts pick the tools and price them. Nothing is sent.
An express draft. For the full one: AI of Common Sense · book a call
Six facts about you
Fill them in and the stack appears below. Name a platform and it gets a second column: the same layers, in that platform's own services, with their own price.
Nothing stored, nothing sent.
Your recommended AI stack
Appears once the brief has who runs it, where data lives, what you are asked for, where you are, your platform.
The whole catalogue
Every tool, by layer and by what kind of thing it is. Once the brief is in, your picks light up here. The stage under each layer is the order from the article: pilot, production, scale.
- open source
- $ vendors
- $$ high price vendor
- first choice in its layer
Model
01Model
Writes the answer: a hosted API, or open weights on an inference server you run.
Stage 1: pilot
hosted API
- Anthropic APIFirst choice in this layer$$ vendors↗Hosted frontier models
- OpenAI APIFirst choice in this layer$$ vendors↗Hosted frontier models
- Amazon Bedrock models$$ vendors↗Models in AWS · AWS
- Azure OpenAI$$ vendors↗OpenAI models in Azure · Azure
- Cortex AI models$$$$ high price vendor↗Models in Snowflake · Snowflake
- Gemini API$$ vendors↗Hosted frontier models · Google Cloud
- Alibaba Cloud Model Studio$$ vendors↗Hosted Qwen
- Cohere$$ vendors↗Enterprise models and rerank
- Databricks Model Serving$$$$ high price vendor↗Inference · Databricks
- DeepSeek API$$ vendors↗Hosted DeepSeek
- Fireworks$$ vendorsModel cloud
- Groq$$ vendorsFast inference
- Mistral$$ vendors↗EU hosted models
- Modal$$ vendorsGPU compute
- Together$$ vendorsModel cloud
- Baseten$$ vendorsModel serving
open weights
- Qwen (Alibaba)First choice in this layeropen source↗Open weights, frontier-class
- DeepSeekopen source↗Open weights, reasoning
- GLM (Zhipu / Z.ai)open source↗Open weights, agents and coding
- Kimi K2 (Moonshot)open source↗Open weights, long context
- gpt-oss (OpenAI)open source↗Open weights from OpenAI
- Gemma (Google)open source↗Open weights, small
- Llama (Meta)open source↗Open weights, the ecosystem default
- MiniMax M-seriesopen source↗Open weights, efficient
- Mistral open weightsopen source↗Open weights, EU
inference server
- vLLMFirst choice in this layeropen sourceInference server
- Ollamaopen sourceLocal models
- SGLangopen sourceInference server
Context
02System prompt
The role and the limits, versioned and tested like code.
Stage 1: pilot
prompt management
- Prompts in gitopen sourcePrompts as code
- PromptLayer$$ vendors↗Prompt management
03Knowledge base
Documents, the data catalogue and the links between them, searched within the user's rights.
Stage 1: pilot
document parsing
retrieval
- pgvectorFirst choice in this layeropen sourcePostgres vectors
- Cortex Search$$$$ high price vendor↗Retrieval in Snowflake · Snowflake
- Databricks Vector Search$$ vendors↗Retrieval in Databricks · Databricks
- Qdrantopen sourceVector database
- Vertex AI Search$$ vendors↗Retrieval in GCP · Google Cloud
- Azure AI Search$$$$ high price vendor↗Enterprise search · Azure
- Bedrock Knowledge Bases$$$$ high price vendor↗Managed RAG · AWS
- BigQuery$$$$ high price vendor↗Knowledge base · Google Cloud
- Chromaopen sourceEmbedded vectors
- Elasticsearchopen sourceSearch engine
- LanceDBopen sourceColumnar vectors
- OpenSearchopen source↗Search engine
- Pinecone$$ vendorsManaged vectors
- Weaviateopen sourceVector database
- Turbopuffer$$ vendorsObject-store vectors
retrieval framework
- LlamaIndexFirst choice in this layeropen source↗Data agents and RAG
- Haystackopen sourcePipeline RAG
data catalogue
04Semantic layer
One definition of each metric and entity, and questions answered over it rather than over raw tables.
Stage 3: scale
metric layer
- CubeFirst choice in this layeropen source↗Semantic layer
- Fabric and Power BI semantic models$$ vendors↗Metrics in Microsoft · Azure
- Snowflake Semantic Views$$$$ high price vendor↗Metrics in the warehouse · Snowflake
- Unity Catalog metric views$$ vendors↗Metrics in Databricks · Databricks
- dbt Semantic Layer$$ vendors↗Metrics on dbt
- Looker (LookML)$$$$ high price vendor↗Metrics in GCP · Google Cloud
- Power BI semantic model$$ vendors↗Metrics in Microsoft · Azure
questions over the model
- AI/BI Genie$$ vendors↗Questions over the lakehouse · Databricks
- Cortex Analyst$$$$ high price vendor↗Questions over semantic views · Snowflake
- Wren AIopen source↗Text-to-SQL on a semantic model
- Amazon Q in QuickSight$$ vendors↗Questions over QuickSight · AWS
- Copilot in Power BI$$ vendors↗Questions over Power BI · Azure
- Gemini in Looker$$$$ high price vendor↗Questions over LookML · Google Cloud
- Vannaopen source↗Text-to-SQL
05Data quality
Says whether the tables behind an answer can be trusted today.
Stage 3: scale
checks as code
Agent
06Harness
Runs the loop: the plan, the calls, the limits, the point where a person signs off.
Stage 1: pilot
agent framework
- LangGraphFirst choice in this layeropen source↗Graph orchestrator
- Microsoft Agent Frameworkopen source↗Enterprise agents · Azure
- OpenAI Agents SDKopen source↗Handoff orchestration
- Pydantic AIopen sourceTyped agents
- CrewAIopen source↗Role crews
- Google ADKopen source↗Graph workflows · Google Cloud
- Gooseopen source↗Local coding agent
- Mastraopen sourceTypeScript agents
- Strands Agentsopen source↗AWS-native SDK · AWS
- Agnoopen sourceLightweight agents
- Hermes Agentopen source↗Reference runtime
- LangChainopen sourceChain toolkit
- OpenClawopen source↗Personal agent gateway
- Microsoft AutoGenopen source↗Group chat · Azure
- Semantic Kernelopen source↗Plugins and processes · Azure
durable workflows
managed agent platform
- Microsoft FoundryFirst choice in this layer$$$$ high price vendor↗Microsoft agent service · Azure
- Amazon Bedrock AgentCore$$$$ high price vendor↗AWS managed agents · AWS
- Databricks Agent Bricks$$$$ high price vendor↗Lakehouse agents · Databricks
- Gemini Enterprise Agent Platform$$$$ high price vendor↗GCP agent platform · Google Cloud
- Snowflake Cortex Agents$$$$ high price vendor↗Warehouse agents · Snowflake
- Salesforce Agentforce$$$$ high price vendorCRM agents · salesforce
- Agynopen source↗Agent runtime control
07Memory
Remembers the conversation and the customer between sessions.
Stage 2: production
08MCP & skills hub
Tools, skills and sandboxes from one catalogue: what the agent can do, and where its code runs.
Stage 1: pilot
MCP gateway and registry
skills
- Agent SkillsFirst choice in this layeropen source↗Skill format
servers over your data
servers over your systems
- AWS MCP serversopen source↗MCP over AWS · AWS
- Azure MCP Serveropen source↗MCP over Azure · Azure
- GitHubopen source↗Code and issues
- Playwrightopen source↗Browser control
- Composio$$ vendorsTool catalogue
- GitLabopen source↗DevSecOps
- Linear$$ vendors↗Issue tracking
- Notionopen source↗Docs and databases
- Sentryopen source↗Error tracking
- Slack$$ vendors↗Team messaging
- Stripeopen source↗Payments
- Vaultopen source↗Secrets MCP
- Nextcloudopen source↗Self-hosted files
09Channel + handoff
Where people talk to it, and how a hard case reaches a person.
Stage 1: pilot
staff chat
- LibreChatFirst choice in this layeropen source↗Multi-provider chat UI
- Microsoft Copilot Studio$$$$ high price vendor↗Low-code agents in M365 · Azure
- Open WebUIopen sourceSelf-host chat UI
- Amazon Q Business$$$$ high price vendor↗Chat in AWS · AWS
- Gemini Enterprise$$$$ high price vendor↗Chat in GCP · Google Cloud
- Snowflake Intelligence$$$$ high price vendor↗Chat in Snowflake · Snowflake
Access
10Identity + access
Acts on the user's behalf or under its own identity, with the right permissions and secrets.
Stage 2: production
on the user's behalf
- Auth0 for AI Agents$$ vendors↗Agent auth
- Descope$$ vendorsAgentic identity
- WorkOS$$ vendorsEnterprise auth
agent identity
fine-grained authorisation
secrets
- OpenBaoFirst choice in this layeropen source↗Secrets
- AWS Secrets Manager$$ vendors↗Secrets in AWS · AWS
- Azure Key Vault$$ vendors↗Secrets in Azure · Azure
- Infisicalopen sourceSecrets platform
- Secret Manager$$ vendors↗Secrets in GCP · Google Cloud
- Databricks secrets$$ vendors↗Secrets in Databricks · Databricks
- Doppler$$ vendors↗Managed secrets
- HashiCorp Vault$$$$ high price vendor↗Secrets
- AWS IAM$$ vendors↗Cloud IAM · AWS
11Filter
Catches prompt injection on the way in and checks answers on the way out.
Stage 1: pilot
injection guard
- Azure AI Content Safety$$ vendors↗Guardrails in Azure · Azure
- Bedrock Guardrails$$ vendors↗Guardrails in AWS · AWS
- Lakera (Check Point)$$$$ high price vendor↗Prompt injection filter
- Model Armor$$ vendors↗Guardrails in GCP · Google Cloud
- NeMo Guardrailsopen source↗Programmable rails
- Cisco AI Defense$$$$ high price vendor↗AI firewall
- Cortex Guard$$$$ high price vendor↗Guardrails in Snowflake · Snowflake
- LLM Guardopen source↗Input and output scanners
- Palo Alto Prisma AIRS$$$$ high price vendor↗AI runtime security
- HiddenLayer$$$$ high price vendorModel security
- Zenity$$$$ high price vendorAgent posture
output validation
- Guardrails AIopen sourceOutput validation
- Llama Guardopen sourceSafety classifier
Run & Cost
12Gateway
One door for every model call: masks personal data before it leaves, keys and budgets, a fallback when a provider fails, a cache that respects rights, routing by price and quality, a log of every call.
Stage 1: pilot
model proxy
- LiteLLMFirst choice in this layeropen source↗Model proxy
- AgentCore Gateway$$ vendors↗Gateway in AWS · AWS
- Azure API Management AI gateway$$ vendors↗Gateway in Azure · Azure
- Mosaic AI Gateway$$ vendors↗Gateway in Databricks · Databricks
- Portkeyopen sourceAI gateway
- Apigee$$$$ high price vendor↗Gateway in GCP · Google Cloud
- Kong AI Gateway$$$$ high price vendorAPI gateway
router
- OpenRouter$$ vendorsModel router
- RouteLLMopen source↗Cost-based routing
cache
- Cloudflare AI Gateway$$ vendorsEdge gateway
Control
13Traces
Records every step the system took.
Stage 1: pilot
14Evals
Tests every change against a set of reference questions.
Stage 1: pilot
CI evals
- Promptfooopen source↗CI evals
- DeepEvalopen sourceTest-style evals
RAG evals
- Ragasopen source↗RAG evals
eval platform
- Bedrock Evaluations and CloudWatch$$ vendors↗Evals in AWS · AWS
- Foundry evaluations and observability$$ vendors↗Evals in Azure · Azure
- MLflow agent evaluation$$ vendors↗Evals in Databricks · Databricks
- Vertex AI Evaluation$$ vendors↗Evals in GCP · Google Cloud
- Braintrust$$ vendorsEval platform
- Cortex AI Observability$$$$ high price vendor↗Evals in Snowflake · Snowflake
15Feedback
Turns bad answers from real users into tests and fixes.
Stage 3: scale
labelling
- Argillaopen source↗Feedback and labelling
Questions
- What is an AI stack?
- The fifteen layers a working AI system is assembled from, in six blocks: model, context, agent, access, cost and control. Most companies need some of them, and the ones you do not need are the expensive part of the answer.
- What does an AI stack cost to run?
- Licences, hosting and people, per month in dollars, for the tools your six facts pick. These are the tools' own prices; building the layers is a separate cost. Naming a platform adds a second column priced in that platform's own services.
- How does this relate to the article on the fifteen layers?
- Same layers, same numbers, same order. The article says what each layer does, what breaks without it and what building it costs; the chooser says which tool to put in it and what that tool costs a month.
- Build or buy?
- Asked per layer, not per stack. The chooser names a tool for each layer you need and dims the ones you do not, so the question only comes up where it changes something.
- Can it run without a cloud?
- Say so in the sixth fact — data may live on our own servers — and the chooser picks from what can run there, with the price that follows.