Public code · self-hosted prototype

Terracotta

A personal AI workspace connecting model selection, tool approval, cost accounting, and self-hosting, with visible control over what gets executed.

ROLEProduct design · UI · agent/API implementation · deployment

PERIOD2026.07–

BUILT WITHTypeScript · React / Next.js · Cloudflare Workers · D1 · MCP · Docker

Source code ↗Self-hosting guide ↗Build history ↗

Terracotta workspace with chat, model selection, and a personal garden — Product interface · model calls require provider API access.

Approval before tool actions: MCP
Container image publishing: GHCR
Persistence and health checks: D1

01 / PRODUCT

Overview

The product starts with the friction of switching AI tools while tracking context and cost separately. Model preferences, connected tools, and work history share a single flow.

02 / SYSTEM

Architecture

Provider model catalogs, price data, and execution usage are kept conceptually separate. Catalog refresh is distinct from a registry that promotes trained model artifacts.

01Router
Task and model preference
02Tool loop
MCP · pending approvals
03Review
A second provider
04Ledger
Usage and execution traces

03 / DELIVERY

Deployment and validation

Docker and Compose configuration, database health checks, backup instructions, and update procedures are published. GitHub Actions builds container images for GHCR with public run records.

The self-hosted edition is a personal prototype. It does not establish multi-user internet operation or high availability.

Image versions and persistent state have separate update and recovery procedures.
CI requires lint, build, and tests before publishing. Fourteen local tests pass, covering D1 restore, encryption keys, provider HTTP 504 fallback, and healthcheck failures.

04 / TRADE-OFF

Evaluation

An offline policy benchmark of 24 tasks estimated lower cost against a high-price single-model baseline, but higher cost against a low-price fixed model. This is not measured savings from live API traffic.

The quality axis depended on hand-defined constants and was excluded from performance claims. The next evaluation compares real classification, task success, and cost together.

ENGINEERING DECISIONS

Engineering decisions

Explicit approval for writes

Tools that change external state expose their service and arguments before execution. Approval state and traces make agent actions inspectable.

Keep state beyond a container

Usage, settings, approval records, and encryption keys persist in a volume. Health checks inspect both the application and database.

Remove an unmeasured quality claim

Hand-defined quality constants in a policy simulation cannot establish model quality. That claim was removed, and cost comparisons retain their baseline conditions.

SCOPE & LIMITS

Limitations and scope

Real model responses require provider APIs. Subscription and credit UX remain prototypes.
Container publication history is distinct from service availability or large-scale model serving.

Source code ↗Self-hosting guide ↗Build history ↗

NEXT CASEMeeting Assistant ↗

Contact

hyunaeee@gmail.com ↗

South Korea · Korean / English