--grant-identity-json <GRANT_IDENTITY_JSON> | JSON file containing a signing-key byte array and a recipient-key codec-hex string |
--stake <STAKE> | Stake amount (in CBY, e.g., 50000). Default: 50000. |
--nonce <NONCE> | Transaction nonce (0 = auto). Default: 0. |
--encryption-pubkey <ENCRYPTION_PUBKEY> | X25519 encryption public key (hex, 32 bytes) for receiving ECDH-sealed volume keys |
--route-serving-endpoint <ROUTE_SERVING_ENDPOINT> | Public base URL for this runner’s CIP-15 off-chain route-serving endpoint |
--statusz-url <STATUSZ_URL> | Operator-advertised /statusz base URL (CIP-42 discovery) |
--job-types <JOB_TYPES> | Job types this runner advertises (comma-separated). Defaults to “llm,http,mcp,agent”. Default: llm http mcp agent. |
--models <MODELS> | CIP-2: model names this runner serves for llm/agent jobs (comma-separated), in the namespace jobs request (e.g. “claude-sonnet-5,gpt-5”). ”*” = any model — the default, written on-chain explicitly; the registry rejects llm/agent runners that advertise no models. Default: *. |
--llm-input-token-price <LLM_INPUT_TOKEN_PRICE> | Input token price in CBY wei per token. Default: 1000. |
--llm-output-token-price <LLM_OUTPUT_TOKEN_PRICE> | Output token price in CBY wei per token. Default: 2000. |
--container-tiers <CONTAINER_TIERS> | CIP-10: isolation tiers the container sandbox provides (comma-separated: standard, strong). Requires “container” in --job-types; advertises all protocol base images and the --container-max-* capacity |
--container-max-cpu-millicores <CPU_MILLICORES> | Largest CPU reservation (millicores) one container job may hold. Default: 2000. |
--container-max-memory-mib <MEMORY_MIB> | Largest memory reservation (MiB) one container job may hold. Default: 2048. |
--container-max-scratch-disk-mib <SCRATCH_DISK_MIB> | Largest scratch-disk reservation (MiB) one container job may hold. Default: 2048. |
--container-max-gpu-count <GPU_COUNT> | GPUs one container job may reserve. Default: 0. |
--serves-registered-images | advertise that this runner provisions + digest-verifies REGISTERED images locally (digest-ref jobs only match runners with this flag). Requires --container-tiers |
--structured-output | CIP-45: advertise that this runner serves structured LLM jobs (runner.llm(..., response_model=...)) through an OpenAI-compatible provider. Requires “llm” in --job-types. A runner whose registration states it without serving them refuses to start |
--inference | Advertise inference.chat_completions in provides, so the chain can assign this runner to inference sessions (cowboy session open --model) |
