Quick start

Install egin, write an agent from a template and run your first prompt.

Install

Create an agent without installing anything first:

npx egin@latest init my-agent

pnpm dlx and bunx work the same way. To keep egin on your PATH instead, install the binary with Cargo:

cargo install egin

To build from source, run cargo build --release in the repository. The binary lands at target/release/egin.

Set one key

Set the API key of the model provider you use, in your shell or in a .env.local file in the directory you run egin from:

export ANTHROPIC_API_KEY=...

That is all the setup there is. egin picks a model for the key it finds:

Key Model
YDC_API_KEY kimi-k3.1-pro on the You.com router, with You.com web search
ANTHROPIC_API_KEY anthropic/claude-sonnet-5-5, searching with Claude’s own web search
OPENAI_API_KEY openai/gpt-6-luna
AI_GATEWAY_API_KEY anthropic/claude-sonnet-5.5 through the Vercel AI Gateway
AWS_BEDROCK_API_KEY us.anthropic.claude-sonnet-5-5 on Amazon Bedrock

Keys can also go in .env.local or .env in the folder you run egin from. A Gemini key alone picks no model: name a gemini/… model in agent.toml.

To use another model, name it in agent.toml. The prefix picks the provider and the key: anthropic/…, openai/…, gemini/…, bedrock/…, and any other vendor/model goes through the AI Gateway. See Providers.

model = "openai/gpt-6.1-sol"

Web search needs no key of its own: egin uses your You.com, Parallel, Firecrawl, context.dev or Brave key, or SEARXNG_URL, if one is set. Otherwise Claude and AI Gateway models use the provider’s own search, and other models go without web tools. egin info my-agent shows what the agent will use and which key, if any, is missing.

Pick a template

egin init writes a starter folder from one of five templates, with agent.toml already naming the model for your key. Without --template, you get basic. --template also takes a path to any folder, to start from your own.

egin init my-agent --template research
Template What it is
basic A minimal agent with one example tool, two evals and two variants
research Answers with sourced facts, writes reports to the workspace, remembers which sources were reliable
ops Follows runbooks, checks hosts on a schedule, blocks destructive commands with middleware
support Answers from a knowledge base over Slack, escalates what it cannot answer
web A Next.js chat app built on shadcn/ui, with the agent in agent/. npm run dev starts both. See Frontends

Run it

egin run my-agent "Who maintains SQLite, and under what license?"

The answer streams to the terminal. The session is stored in my-agent/.egin/egin.db, so you can continue it later:

egin run my-agent --session <id> "And when was it first released?"

By default egin asks before running a tool that is not marked safe. Pass --yes to approve everything for one run, or set approval = "never" in agent.toml.

Check it

Every template ships with evals. Run them:

egin eval my-agent

Each eval is a prompt plus a check. The run prints a pass or fail per task and writes results to my-agent/evals/results/. See Evals for the format.

Serve it

egin serve my-agent --port 8787

Now any program can start sessions over HTTP and follow them over SSE. See the HTTP API.