Skip to content

SKILL.md

Latest commit

 

History

History
107 lines (85 loc) · 4.34 KB

SKILL.md

File metadata and controls

107 lines (85 loc) · 4.34 KB
Raw
name setup-omnara-agent
description Install and set up Omnara. Use when the user asks to install Omnara or run /setup-omnara-agent.

Set up an Omnara agent

Get the user a launched Omnara agent and its link. The commands below are complete; read other docs only if one fails or the user wants Slack, Discord, GitHub, or their own app.

If the Omnara MCP server is attached (tools such as whoami, profiles_create, and agents_launch), use the Omnara MCP tools for every step they cover. Do not install or log in to the CLI for those steps. Call whoami; if it fails with an authorization error, ask the user to complete the Omnara browser authorization for the MCP server in their client, then call whoami again. Do not continue setup until it succeeds.

1. Sign in

npx omnara whoami

If it says you're not logged in, start npx omnara login without piping its output, in the background if you can, and keep it running while the user approves. It opens the approval page in the user's browser and prints a link and a code. Before you wait for it, send the user a message with both, since their browser may be on another machine, and ask them to approve once the code matches. Do not continue setup until login completes and npx omnara whoami succeeds.

2. Ask what to build

Ask two questions before creating anything:

  1. What should the agent do? Examples include a coding agent, deep research agent, etc.
  2. How will you interact with the agent (Omnara web console, Slack, GitHub, Discord, or your own app using the public API/TypeScript SDK)?

3. Pick the project, model, and pool

Look up the user's org and its default project (npx omnara projects list --org <org-id> --json lists projects). Then:

npx omnara config --org <org-id> --project <project-id> &&
npx omnara grant models list --json | node -e 'for (const g of JSON.parse(require("fs").readFileSync(0)).data) console.log(g.model.provider_config, g.model.name)' &&
npx omnara grant pools list --json | node -e 'for (const g of JSON.parse(require("fs").readFileSync(0)).data) console.log(g.machine_pool.name)'

Use a listed model, and default-pool if it's listed, otherwise the first pool.

4. Create the agent and launch it

Fill in an instruction for what the user asked for, a short profile name, and a simple first message. The machine tools come with machine_sources; don't list them, or other built-in tools the user doesn't need.

Add relevant secrets or startup scripts for the machine pool env override. For example, if the user wants to clone a GitHub repository, you may set up a script (machine_provider_options_overlay.startup_script on the machine source) which clones the repo upon starting the machine. If needed, you can pipe a GitHub PAT via a secret into the env var overlay.

Then create the profile and launch the agent:

cat > omnara-agent.yaml <<'EOF'
version: v1
instruction: |
  <what the agent does, for whom, and how it reports back>
model:
  provider_config: <provider>
  name: <model>
tools:
  web_search: {}
  web_fetch: {}
  ask_question: {}
machine_sources:
  - machine_pool_name: <pool>
EOF
npx omnara profiles create --name <profile-name> --file omnara-agent.yaml --json > omnara-profile.json &&
npx omnara agents launch --profile "$(node -p 'require("./omnara-profile.json").id')" --message "<first message>" --json | node -e 'const r=JSON.parse(require("fs").readFileSync(0)); console.log(r.agent.id)'

Give the user the link: https://app.omnara.com/projects/<project-id>/agents/<agent-id>.

5. Other places to talk to it

Only if the user chose them: