DeesseJS Agents

A monorepo of specialized AI agents, each shipped as an independent Vercel project. Every agent plays a "head of" role (engineering, product, fitness, finance, ...) and shares one Postgres database for cross-agent memory. Maintained and used internally by the deessejs.com teams.

Repository layout

ds-team/
├── apps/                       # one folder per agent = one Vercel project
│   ├── head-of-engineering/
│   ├── fitness-coach/
│   └── ...
├── packages/
│   ├── agent-core/             # shared channels (eve, telegram) + MCP connections (exa, github, ...)
│   ├── database/               # Drizzle + Neon; shared memories + memory_audit
│   └── typescript-config/      # shared tsconfig presets
├── docs/
│   ├── internal/agents/        # per-agent docs, grouped by status (idea/design/scaffolded/deployed/retired)
│   ├── internal/product/       # product specs
│   └── internal/reports/       # design docs (memory schema, MCP integrations, skills)
├── turbo.json
└── pnpm-workspace.yaml

Stack

• Agent framework: eve v0.20 (runs on Vercel)
• Model SDK: Vercel AI SDK ai v7
• Model: MiniMax-M3 via vercel-minimax-ai-provider (direct provider, not the AI Gateway)
• Auth: @vercel/connect (OIDC)
• Database: Postgres on Neon, accessed via Drizzle ORM (@neondatabase/serverless)
• MCP tools: Exa (web search/fetch), GitHub (repos, issues, PRs), Liftosaur (fitness), Todoist (in progress)
• Channels: eve HTTP, Telegram
• Build orchestrator: Turborepo
• Package manager: pnpm workspaces
• Runtime: Node 24.x, TypeScript 7.0.1-rc

Prerequisites

• Node.js 24.x
• pnpm 9.15+
• A Vercel account (each apps/<agent>/ becomes its own Vercel project)
• API keys for the providers your agents use (see .env.example)

Setup

# enable pnpm if you don't already have it
corepack enable

# install dependencies for the whole workspace
pnpm install

# copy the env template and fill in real values
cp .env.example .env

Each agent also ships its own .env.example under apps/<agent>/ with the variables it actually needs (most inherit from the root .env).

Commands

Run from the repo root. Turborepo fans out across all packages.

Command	What it does
pnpm dev	Start every agent's dev server in parallel
pnpm build	Build every agent
pnpm typecheck	Run tsc --noEmit everywhere
pnpm start	Run the production build of every agent

To target a single app:

pnpm --filter fitness-coach dev
pnpm --filter head-of-engineering build

Database

The shared memories and memory_audit tables live in packages/database and are migrated through Drizzle Kit. Agents read and write each other's memories through the tool factory in packages/database/src/tools/memory.ts, which closes over the calling agentId so memories are scoped correctly.

# generate a migration after editing schema.ts
pnpm --filter @ds-team/database generate

# apply pending migrations
pnpm --filter @ds-team/database migrate

# open Drizzle Studio
pnpm --filter @ds-team/database studio

Deploying an agent

1. In the Vercel dashboard, Add New → Project and import this repo.
2. Set Root Directory to the chosen app (e.g. apps/fitness-coach).
3. Add the project's environment variables (or migrate to Vercel Connect).
4. Deploy.

Pushes that don't touch an app are automatically skipped (Turborepo's content-aware hashing), so shared-package changes don't redeploy every agent.

Adding a new agent

The agent lifecycle is idea → design → scaffolded → deployed. Short version:

1. Drop a one-pager describing the idea, scope, and capabilities.
2. When the design firms up, write a design doc with a dated filename.
3. Scaffold apps/<id>/, register the agentId in KNOWN_AGENT_IDS (packages/database/src/schema.ts), and run pnpm --filter @ds-team/database guardrail.
4. Provision the Vercel project (and a Telegram bot, if needed). Update the registry.

Conventions

• Naming: kebab-case everywhere (folder name, agentId, package name, Vercel project name, Telegram bot handle).
• Shared infra is reused, never forked: channels and connections come from @ds-team/agent-core, persistence from @ds-team/database.
• MCP surfaces default to permissive: no client-side filtering, no read-only mode, no approval gate. Behavior is shaped by the agent's instructions.md, the canonical skill bodies, and per-tool HITL where the channel offers it.
• Memory is always written through @ds-team/database with the agentId baked into the tool closure. Direct table access is discouraged.

License

MIT. See LICENSE for the full text.