Quickstart

• Linux on amd64 or arm64, or macOS on Apple silicon.
• Docker, because local Runtime opens Task Environments as containers.
• uv, which installs fleet-runtime in an isolated tool environment.
• gh, authenticated to a GitHub account that can read the private fleet-ai/platform repository.
• A model-provider key when the agent harness you plan to use requires one.

export PATH="$HOME/.local/bin:$PATH"
(
  set -euo pipefail
  CLI_VERSION=cli-v0.1.1
  case "$(uname -s)-$(uname -m)" in
    Linux-x86_64) flt_asset=flt-linux-amd64 ;;
    Linux-aarch64 | Linux-arm64) flt_asset=flt-linux-arm64 ;;
    Darwin-arm64) flt_asset=flt-darwin-arm64 ;;
    *) echo "Unsupported platform: $(uname -s)-$(uname -m)" >&2; exit 1 ;;
  esac

  release_dir=$(mktemp -d)
  trap 'rm -rf "$release_dir"' EXIT
  gh release download "$CLI_VERSION" --repo fleet-ai/platform --dir "$release_dir"
  if command -v sha256sum >/dev/null; then
    (cd "$release_dir" && sha256sum -c SHA256SUMS)
  elif command -v shasum >/dev/null; then
    (cd "$release_dir" && shasum -a 256 -c SHA256SUMS)
  else
    echo "A SHA-256 checksum utility is required." >&2
    exit 1
  fi

  mkdir -p "$HOME/.local/bin"
  install -m 0755 "$release_dir/$flt_asset" "$HOME/.local/bin/flt"
  uv tool install --force \
    --constraints "$release_dir/fleet-runtime-constraints.txt" \
    "$release_dir"/fleet_runtime-*.whl

  flt --version
  fleet-runtime --version
  flt config set registry registry-alpha.fleetai.me
  flt auth login registry-alpha.fleetai.me
)

These commands select the released flt binary for Linux amd64, Linux arm64, or Apple silicon macOS, then download one pinned, immutable CLI release. Change CLI_VERSION to install another published release. uv tool installkeeps Runtime and its dependencies in a dedicated environment. The checksum is verified before installation, and --force makes the process safe to repeat. Run uv tool update-shellif ~/.local/bin is not already on your persistent shell path. The config command selects the Registry used by shorthand and platform commands. Login prints a Fleet SSO URL and stores the credential both CLIs use.

export ANTHROPIC_API_KEY="..."

run_id=$(fleet-runtime runs submit \
  --prompt "Use bash to print hello, then submit hello" -n 1 \
  --agent claude --harness-var ANTHROPIC_API_KEY=@ --follow)

fleet-runtime runs show "$run_id"

A prompt run creates a real one-Task TaskSet and a durable Run, but it is intentionally ungraded because no Verifier defines a right answer. Use an authored TaskSet that declares the required Verifier or Verifiers when the result must carry a score or verdict.

1. 1
   Open Tasksets

   Choose a repository and an exact version.

2. 2
   Select New experiment

   Choose all Tasks, a sample, or exact Task keys.

3. 3
   Configure an agent

   Pick the harness, model, rollout count, result project, and credential sources.

4. 4
   Create the Experiment

   Fleet freezes the selection and creates one Run per agent configuration.

A project is the ownership and access scope that contains Repositories. If you need a destination for your own results, open Projects and choose Create project. You become its owner. Project grants can provide read, write, or admin access across the project, while Repository grants provide read or write access to one Repository. A project owner or administrator can also share a Repository capability invitation link for you to accept.

Open Tasksets and search by project, Repository name, short digest, or Repository-label key or value. Visibility is the only structured TaskSet filter. Labels can help locate an imported collection, but they are metadata—not an authoritative “Fleet-authored” provenance filter.

Creating an evaluation needs read access to the source Repository, write access to the results project, and read access to each selected project credential. See Access required for hosted evaluations.