Quickstart

Get Elliot running and talking to an agent in minutes.

Run Elliot (Docker — recommended)

The fastest way. The only prerequisite is Docker — no Python, Node, uv, or pnpm, and no source checkout:

curl -LsSf https://raw.githubusercontent.com/EliBarak12/Elliot/main/scripts/install.sh | sh

This pulls the pre-built images, generates a local .env with a fresh secret key, starts all three services, and opens Studio at http://localhost:8080.

# stop
docker compose -f docker-compose.run.yml down
# view logs
docker compose -f docker-compose.run.yml logs -f

On the Docker path you build and test connectors visually in Studio. To use the elliot CLI (init, lint, eval), run from source instead.

Run from source

For developing Elliot itself, or to use the elliot CLI.

Prerequisites

• uv (Python 3.13)
• pnpm (Node 22)

curl -LsSf https://astral.sh/uv/install.sh | sh
npm install -g pnpm

Clone & install

git clone https://github.com/EliBarak12/Elliot.git
cd Elliot
make setup
cp .env.example .env

Boot the stack

make dev

make dev runs elliot connect first — auto-registering Elliot with every coding agent it can find (Claude Code, Cursor, OpenClaw, Codex). Then it brings up:

• elliot-mcp-plugin on :3000 (MCP endpoint)
• elliot-connector-runtime on :3001 (tool execution)
• elliot-studio on :5173 (dashboard)

Open http://localhost:5173.

Scaffold your first connector

elliot init --template rest-api-key my-api.connector.json

Open the file. Fill in the source URL, the auth secret name (an env var), and any tools you want exposed. See Connector spec for the full schema.

Lint & eval before shipping

elliot lint my-api.connector.json
elliot eval my-api.eval.yaml

The linter checks every tool against the five principles. The eval harness is deterministic — no LLM is involved. Each case names a tool and the arguments to call it with; Elliot executes that tool directly against the connector and asserts on the result (row counts, fields present, token size, error codes), reporting a pass/fail with token counts. See the CLI reference for the suite schema.

Call a tool from an agent

In Claude Code (or any registered MCP client), ask it a question that needs your data. The agent calls list_animals (or whatever you defined), Elliot returns a small, contextually-sized response, and Studio shows the call in the audit log.

What an agent gets on first connect

On first connect the agent automatically calls prompts/get name=getting_started. That single prompt teaches the agent the five principles, the canonical workflow (discover-source → build-connector → lint-connector → run-eval → deploy), and the reference resources available (templates, error-code dictionary, install docs).

Next steps

• Concepts — sources, tools, skills, connectors
• The five principles — the design rules
• Architecture — how the three services fit together
• Deployment — Docker images, env vars, hosting