--- name: aither-awdk-quickstart description: >- Install awdk and run an AI agent on your own machine, then find, inspect and install packs and skills with the adk CLI. Covers `adk onboard --quick`, `adk doctor`, the pack verbs (list, search, info, install, enable), the skills verbs (list, search, install, export), `adk tools list`, and running an agent locally with `adk run`. Every command here was run against the installed awdk CLI; flags that are not in each command's `--help` are not in this skill. --- # aither-awdk-quickstart — install awdk, run an agent, install a pack [awdk](https://github.com/Aitherium/awdk) (MIT, on PyPI as `awdk`) is a Python agent runtime: the agent loop, a tool registry, packs of tools, a skills installer and a CLI called `adk`. The model can be local (llama.cpp, vLLM, Bonsai) or any hosted provider you already have a key for. **Check:** `adk --version` prints a version. ## 1. Install ```bash pip install awdk adk --version ``` `adk --help` lists the subcommand surface. Python 3.10 or newer. ## 2. Bring up inference and verify the install ```bash adk quickstart-local --dry-run # print what it would detect and pull; changes nothing adk quickstart-local # detect CPU/RAM/GPU, pick a backend, pull and serve a model adk doctor # toolkit, packs, inference, enrollment — one report ``` `adk quickstart-local` writes its endpoint on a local port (default `8209`). Backends: `--backend auto|llamacpp|vllm|bonsai`. To use a hosted provider instead, set a key — the CLI prompts, so the key never lands in your shell history: ```bash adk keys set anthropic # or openai / deepseek / openrouter / groq / together / google adk models # what this install can actually reach ``` One-shot setup instead of the two steps above: ```bash adk onboard --quick # detect hardware, run inference, install the default pack, enroll adk onboard --quick --pack hermes ``` **Check:** `adk doctor` reports inference reachable and at least one pack installed. ## 3. Find and install packs A pack is a bundle of tools an agent can call. Four verbs cover the whole lifecycle: ```bash adk pack list # installed + available adk pack search "pdf" # search by name, description or tags; no query = everything adk pack info # what it installs, its tools, its trust tier adk pack install # a bare id resolves across the pack shelves adk pack enable # turn it on for this agent (installs first if missing) adk pack update # or with no id: every installed pack adk pack remove ``` A bare id resolves in order: first-party agent packs, marketplace/library packs, then gateway extensions. Pin the shelf explicitly when a name is ambiguous: `awpack:`, `library:`, `ext:`, `skill:`. A community pack nobody at Aitherium reviewed refuses unless you pass `--allow-community` — read it first. `adk pack install` also takes a local bundle: `adk pack install ./dist/mypack-0.1.0.tar.gz`. The community marketplace path is `community:`. **Check:** `adk pack info ` shows it, and `adk tools list` lists its tools. ## 4. Find and install skills Skills are instruction documents (`SKILL.md`), not code. The built-in shelf is the public awskills pack; `install` copies one into a harness skills directory. ```bash adk skills list # learned + installed adk skills search "git" # learned skills + the awskills shelf adk skills install acp-drive --target claude # ~/.claude/skills (Claude Code and awsh) adk skills install aither-interop --target project # ./.claude/skills, this repo only adk skills export # every skill as agentskills.io JSON on stdout adk skills update # re-install what is already installed adk skills remove ``` Third-party skills come from the two places Hermes Agent reads, fetched as a single `SKILL.md`: `hermes://` (a tap, `/skills//SKILL.md`) and `agentskills://`. Both are community trust unless the owner is Aitherium, so they need `--allow-community`. **Check:** `adk skills list` shows the skill under the target you chose. ## 5. Run an agent locally ```bash adk install pack:openclaw # or pack:hermes / pack:claude-code adk run --agents openclaw # start the agent server adk agents ls # agent names you can talk to adk chat # talk to a running agent by name ``` `adk run` flags that matter: `-m/--model`, `-b/--backend`, `-p/--port`, `-a/--agents` (comma-separated identities), `--mesh` (advertise tools and inference to connected peers). To run a persistent agent tied to a fleet, use `adk up` instead. **Check:** `adk run` stays up and `adk chat ` answers a prompt. ## Exit codes Every `adk` verb exits `0` clean, `1` refused or invalid, `2` could not judge. In a script, treat `2` as a failure — it is not a pass.