Skip to main content

Command Line

The ACI CLI

Use the ACI CLI when the workflow belongs in a terminal: local serving, shared-service inference operations, native edge-runtime packaging, or MCP startup.

Local API serving

Run ACI Inference locally with `aci serve` for development, test harnesses, and packaging validation.

Structured-task cloud starting point

For current structured-task cloud deployments, create tenants and keep memory off in the first baseline.

Thin inference wrapper

The inference subcommands call the same HTTP contract exposed by `ACIInferenceClient` and the API docs surface.

JSON-first workflows

Mutation verbs accept payloads from `--json`, `--file`, or `--stdin`, which keeps scripts and CI paths explicit.

Runtime and MCP helpers

The same binary also exposes native runtime packaging helpers and stdio MCP startup.

Install and entry points

Installing the Python package provides `aci`, `aci-mcp`, and `aci-inference-mcp`. The CLI itself is the operator and integration wrapper around the documented public surfaces.

pip install aci-engine

aci --version
aci serve --host 127.0.0.1 --port 8000
aci-mcp
aci-inference-mcp

Environment defaults

ACI_BASE_URL

Default base URL for inference and MCP flows.

ACI_API_KEY

Default API key for inference and MCP flows.

ACI_TIMEOUT_SECS

Default request timeout in seconds.

ACI_MCP_ALLOW_ADMIN

Default admin-tool toggle for MCP startup.

What the CLI is for in delivery

The CLI is not a separate product. It is the operator and packaging surface that comes with the wheel, manages the service locally, and exports the native edge artifacts.

Operator install

The CLI arrives with the `aci-engine` wheel and gives operators one entry point for local service startup, health checks, and tenant lifecycle work.

Edge packaging

Use the CLI as the build and export surface for native runtime artifacts; the final shipped edge runtime is the compiled package it produces, not the Python CLI itself.

MCP startup

The CLI also starts the stdio MCP server, but that process still sits on top of an existing ACI Inference deployment rather than replacing it.

Serve the API locally

`aci serve` launches the FastAPI app locally. It is the fastest path for contract testing, local integration work, and MCP attachment against a development instance.

aci serve --host 0.0.0.0 --port 8000 --log-level info

curl http://127.0.0.1:8000/livez
curl http://127.0.0.1:8000/readyz

Inference command families

Inference subcommands mirror the public HTTP surface: model registration, tenant lifecycle, operational verbs, and health checks. The legacy `aci platform ...` path remains supported as an alias.

Models and tenants

aci inference models list --base-url "$ACI_BASE_URL" --api-key "$ACI_API_KEY"

aci inference models register \
  --base-url "$ACI_BASE_URL" \
  --api-key "$ACI_API_KEY" \
  --model-id support-assistant \
  --device cpu

aci inference tenants create \
  --base-url "$ACI_BASE_URL" \
  --api-key "$ACI_API_KEY" \
  --model-id support-assistant \
  --tenant-id tenant-a \
  --target-dim 4

Verbs and audits

aci inference bind --base-url "$ACI_BASE_URL" --api-key "$ACI_API_KEY" --file bind.json

aci inference infer \
  --base-url "$ACI_BASE_URL" \
  --api-key "$ACI_API_KEY" \
  --json '{"model_id":"support-assistant","tenant_id":"tenant-a","namespace":"default","input":"hello","use_memory":false}'

aci inference audit \
  --base-url "$ACI_BASE_URL" \
  --api-key "$ACI_API_KEY" \
  --model-id support-assistant \
  --tenant-id tenant-a

aci inference livez --base-url "$ACI_BASE_URL"
aci inference readyz --base-url "$ACI_BASE_URL"

Payload sources

The mutation commands `bind`, `adapt`, `constrain`, `infer`, `unbind`, and `consolidate` require exactly one payload source.

# Inline JSON
aci inference constrain --api-key "$ACI_API_KEY" --json '{"model_id":"m1", ... }'

# JSON file
aci inference bind --api-key "$ACI_API_KEY" --file bind.json

# STDIN
cat infer.json | aci inference infer --api-key "$ACI_API_KEY" --stdin

Runtime and MCP helpers

The CLI also exposes native build helpers and the stdio MCP launcher, so packaging and agent attachment stay inside the same public binary surface.

aci edge-runtime build-native --output-dir build/native
aci edge-runtime package-embedder --output-dir build/embedder

aci mcp serve \
  --base-url "$ACI_BASE_URL" \
  --api-key "$ACI_API_KEY" \
  --allow-admin

# equivalent branded entry point
aci-inference-mcp

Exit behavior

  • Successful commands return `0`.
  • Inference client errors print a structured JSON error payload to stderr and return `1`.
  • Keyboard interruption returns `130`.

Need typed programmatic access? Use the Python SDK. Need agent tool-use over stdio? Use the MCP server.

The CLI does not expose private implementation details beyond the shipped public commands, and the legacy `platform` and `runtime` command names remain available as aliases.

SDK MCPAll docs