Read before write
The first script the agent runs only inventories the machine: disks, mounts, containers, ports. Nothing is created or changed until you have read the results and approved a plan.
Vibe integration
Vibe coding writes an app.
Vibe integration assembles a stack.
Market research needs a tick store, a history store, an analytical engine, object storage, charts and a dashboard. No single product is best at all of them. Vibe integration takes the best tool for each job, joins them through small, tested contracts, and hands the installing, wiring and upkeep to an AI agent that works from a runbook you approve.
Best of breed
Each job goes to the tool that does it best.
Integrated suites trade depth for convenience. We keep the depth and let the agent absorb the convenience cost: more moving parts, but each one is the specialist.
Research queriesDuckDB + DuckLake
An in-process SQL engine with a catalog that grows from one file to PostgreSQL.
Control and catalogPostgreSQL
Query library, job leases and the shared catalog for multi-user deployments.
Data and widgetsOpenBB (now OpenBQ)
The backends and widget protocol BDOBB reads, so existing widgets keep working.
The seams
Cohesion comes from contracts, not from one vendor.
A stack of specialists only feels like one product if every join is narrow, documented and tested. These are the seams our three tools are built around.
| Seam | Joins | Contract |
|---|---|---|
| Widget protocol | OpenBB backends → BDOBB | The OpenBB Workspace widget contract, read as-is |
| Live tick stream | kdb+ → Art’s Charts | WebSocket series from the tick cache |
| Completed periods | Delta tick vault → charts and queries | One Delta table per symbol and date |
| Study catalog | Polars registry → chart client | GET /ta_registry; the client catalog is generated from it |
| Research catalog | DuckLake → TemporalDB queries | SQL catalog plus Parquet files Design |
| Agent tools | Model → Art’s Charts | MCP tools for chart creation, snapshots and analysis In testing |
Agent-run install
The agent does the work.
You hold the approvals.
Step through how an agent brings up the stack on a home server. Every change waits at a gate for your OK. Commands are illustrative.
Guardrails
What makes an agent-maintained stack safe to trust.
Pinned versions
Every part runs at a known version, such as Lightweight Charts 5.0.8 and tagged container images, so an upgrade is a deliberate, reviewable change, not drift.
Rollout, then acceptance
Each change ships in two steps: roll it out, then run acceptance checks against the live service. A change that fails acceptance is rolled back, not patched in place.
Recovery is rehearsed
Maintenance includes drills: restart the tick cache and replay its log, restore a catalog to a scratch environment. See the Live cache page for the kdb+ procedure.
Secrets stay out of logs
API keys go into the services that need them and nowhere else. TemporalDB’s design keeps credentials out of every stored request record.
Your model, your choice
The runbooks are plain text and shell, so any capable coding agent can follow them, whether it runs on a hosted model or on a local one on your own hardware.
Companion series
Adventures in OpenBQ, on Substack.
The companion series explains how we got here: why each part was chosen and how the pieces were joined. Every installment comes with instructions you can run by hand, or hand to an AI coding agent such as Claude Code or Codex and approve step by step.
How we got here
The decisions behind the stack, in the order we made them, and the reasoning behind each one.
By hand or by agent
The same instructions work both ways: paste the commands yourself, or give them to your agent as its runbook and keep the approvals.
The gotchas, up front
Each installment lists the gotchas we hit, so you don’t spend an afternoon debugging a problem we have already solved.
Status
How far along this is.
This is how we build and run our own stack today. Turning it into a kit anyone can hand to their agent is the next step.
| Agent-written runbooks and scripts, reviewed before each run | In use |
| Read-only inventory before any change | In use |
| Separate rollout and acceptance steps for chart services | In use |
| MCP tools so agents can drive Art’s Charts | In testing |
| Packaged install kit an agent can run end to end | Planned |
| Scheduled maintenance: upgrades and recovery drills | Planned |
The best tools, without the integration tax.
You pick the specialists. The agent does the plumbing. You stay in charge of every change.