create-evestack

One command to a self-hosted eve agent: durable Postgres sessions, a Docker sandbox, long-term memory, and a dashboard you can actually drive. No Vercel account. $0 infrastructure.

npx create-evestack my-agent

That can be the whole setup. The scaffolder asks four questions up front — where, which model, whether you want Composio's tools, and whether it should bring the stack up for you — and then works while you walk away:

1. Where — the project directory.
2. Model — OpenAI, Anthropic or Ollama, and a key if you have one.
3. Tools — Composio's one-click tool sign-in, on by default.
4. Bring it up — start Postgres, create the schema, pull the dashboard.

Say yes to question 4 and it does all three, then offers to start the agent too, so there is nothing left to paste. Say no — or pass --yes, or leave Docker stopped — and it prints these four commands when it finishes instead:

cd my-agent
docker compose up -d postgres              # durable sessions
npm run db:bootstrap                       # create the workflow schema
docker compose --profile dashboard up -d   # the dashboard, usually on :4000
npm run dev                                # the agent — holds this terminal

The order matters: npm run dev is a foreground process that holds the terminal until you Ctrl-C it, so it goes last. The dashboard port is whichever one was free when you scaffolded, 4000 unless something already had it; .env.local and docker-compose.yml carry the real number, and npx evestack open prints it.

Either way you get a .env.local with a freshly generated auth password, and dependencies installed.

The only thing that costs money is model tokens.

Two names, one scaffolder

npx evestack create my-agent runs this exact code. evestack is the single command, and it answers to eight verbs — create, status, tour, open, verify, skills, attach, doctor — of which create and attach are routed straight into this package's own modules. create-evestack is the name npm's create-* convention leads people to, and it keeps working. Same prompts, same flags, one place a bug gets fixed.

This package is dependency-free on purpose, which is why the implementation lives here rather than the other way round: evestack doctor needs a Postgres driver, and nobody should download one to scaffold a project.

What you get

A project containing:

| | | | --- | --- | | agent/ | the eve agent — instructions, tools, channels, sandbox config | | lib/memory.ts | semantic long-term memory on your own Postgres via pgvector | | evals/ | a runnable eval suite (npx eve eval) | | docker-compose.yml | the pgvector Postgres that stores durable sessions — safe to commit, because the password is not in it | | .env.local | generated, gitignored, with a unique EVESTACK_AUTH_PASSWORD | | .env | generated, gitignored, holding EVESTACK_DB_PASSWORD — the one variable docker-compose.yml interpolates. Compose reads .env and never .env.local, which is the whole reason there are two files. Rotate it in both together: Postgres only applies a new password when the data volume is first created |

and, wired up for you:

• Durable sessions on your Postgres via @workflow/world-postgres — restarting the process loses nothing, because the state never lived in the process.
• A local Docker sandbox instead of hosted Vercel Sandbox.
• HTTP Basic route auth, because eve's stock vercelOidc() / placeholderAuth() are both wrong off Vercel and fail closed.
• Sign-in to 1,000+ tools through Composio, the managed-OAuth ones in one click, if you supply a key. Skipped cleanly if you don't.
• Trace export to the evestack dashboard, if EVESTACK_DASHBOARD_URL is set.

Requirements

• Node 24+
• Docker, running — Postgres and the agent sandbox both need it
• A model API key, or Ollama for a genuinely $0 stack. Check your free RAM before choosing Ollama: budget roughly both model sizes + 4 GB — long-term memory needs the chat model and a separate 274 MB embedding model loaded, not one of them.

Model providers

The scaffolder asks which one you want and writes both variables to .env.local. EVESTACK_PROVIDER is the one agent/agent.ts branches on — a model name written without it goes to whichever provider was already selected, so change them together:

| EVESTACK_PROVIDER | Key | Default EVESTACK_MODEL | | --- | --- | --- | | unset, or openai | OPENAI_API_KEY | gpt-5-mini | | anthropic | ANTHROPIC_API_KEY | claude-sonnet-5 | | openrouter | OPENROUTER_API_KEY | qwen/qwen3.8-27b | | ollama | none | qwen3:0.6b | | compatible | EVESTACK_COMPATIBLE_API_KEY (optional) | none — plus EVESTACK_BASE_URL |

openrouter is one key for 400+ models, open weights included. compatible is anything else speaking the OpenAI wire format — LM Studio, llama.cpp, vLLM, Groq, Together.

The dashboard

The agent is only half of evestack. The other half is an open replacement for Vercel's Agent Runs — sessions, turns, subagent trees, computed cost per turn, plus the ability to start sessions, stream replies and resolve approvals.

It is already in the docker-compose.yml this scaffolder writes, behind a profile so a plain docker compose up -d does not pull ~230 MB on someone who only asked for a database:

docker compose --profile dashboard up -d

Nothing to clone and nothing to build: the service pulls ghcr.io/sammytourani/evestack-dashboard, reads the same .env.local your agent reads, and reaches the same Postgres. Sign in with the EVESTACK_AUTH_USER / EVESTACK_AUTH_PASSWORD the scaffolder generated — it prints them when it finishes.

Set EVESTACK_DASHBOARD_IMAGE to run an image of your own instead.

Documentation

Full docs, architecture notes and troubleshooting live in the evestack repository.

License

Apache-2.0, same as eve. eve is a trademark of Vercel; evestack is an independent project and is not affiliated with or endorsed by Vercel.