MetaMCP, Airtable MCP, and Airtable AI Agent show the interface layer: tool surfaces agents can call.
Why this fits
Cotizera and data warehouse work show the operating layer: records, PDFs, follow-up, and data models.
The cloud path is concrete: Google ADK, Agent Registry, Cloud Run, IAM, Secret Manager, Docker, and CI/CD.
Relevant work
Repos and systems, not claims.
The useful signal is the pattern: tools, records, permissions, deployment, and evidence.
MCP gatewayMetaMCP
A gateway for turning many MCP tools into a smaller surface that can run beyond a local desktop.
Repo
Tool contractAirtable MCP
Airtable exposed as MCP tools: bases, tables, records, schema, reads, writes, and natural language access.
Repo
Python agentAirtable AI Agent
A Python agent that turns Airtable operations into tool calls, workflows, and repeatable actions.
Repo
Sales workflowCotizera Agents
Quote intake, PDF generation, WhatsApp follow-up, and pipeline updates shaped as agent-assisted work.
Case study
Evidence layerMentu Protocol
Commitments, ledgers, hashes, and evidence for work that needs to be reviewed after agents run.
Repo
Business dataData warehouse work
Sync bridges, Postgres, reporting layers, and business data models that give agents context.
Case studyHow I work
The hard part is around the model.
An agent can call a tool. It still needs business meaning, permissions, and records it can trust.
MCP, APIs, contracts, and tool calling turn software into a surface an agent can use.
Prompt regression, observability, CI/CD, logs, and ledgers make agent behavior reviewable.
Production pattern
What I would ship.
A useful agent is not one thing. It is a chain of ordinary parts that keep the model honest.
Vertex AI/Gemini provides reasoning. Google ADK gives the agent shape.
Agent Registry gives agents, MCP servers, tools, and endpoints a catalog.
Cloud Run, Docker, IAM, Secret Manager, and CI/CD turn the agent into something deployable.
MCP and APIs connect the agent to Airtable, Cotizera, data warehouses, and internal systems.
Prompt regression, observability, logs, and evidence make the work legible.
Writing
The thinking behind the code.
The agent work sits on top of a larger idea: business data needs meaning before AI can act on it.
Data contextThe Data Infrastructure Nobody Wants to Build
Why AI analytics breaks when the database has facts but not business meaning.
Read
Analysis workflowHow AI Should Analyze Your Data
Why good AI analysis needs a process, not a magic answer box.
ReadAgent designThe Intelligent Sales Agent
How a sales agent emerges from calls, transcripts, scripts, KPIs, and structured memory.
ReadMCP and APIsGetting Software to Talk to Each Other
The bridge between business software, MCP adapters, and AI-readable tools.
ReadCase studies
Let the deep dives carry the detail.
The landing page should orient. These pages hold the evidence and images.
Stack I can work inside
PythonGCPGoogle ADKAgent RegistryVertex AI/GeminiCloud RunDockerCI/CDAzure DevOpsMCPtool callingprompt regressionobservabilityIAMSecret Manager
GCP sources
The platform details have sources.
The GCP side of this work follows the official docs for Agent Registry, external MCP servers, Cloud Run MCP hosting, and ADK toolset consumption.
Through line
The through line is operational AI.
Not a demo persona. A working pattern: data with meaning, tools with contracts, agents with deployment paths, and evidence after important actions.