Run agents locally

Local execution keeps the agent workflow on your machine. Choose a local model or a compatible remote model endpoint, and decide which connections the run may use.

Standalone development

The standalone agentkit workflow command runs a local directory without an Adios account. Use recorded model and connector outputs for repeatable tests, or supply --model-url and --model for a real model.

./dist/agentkit workflow --workflow hello \
  --model-url http://localhost:1234/v1 --model YOUR_MODEL \
  --model-key-env AGENT_MODEL_KEY \
  --input ./hello-workflow/tests/input.json \
  --state ./.agentkit/local --run-id local-hello \
  ./hello-workflow

Set AGENT_MODEL_KEY in your shell or secret manager before running this example. Omit the flag for a server that needs no authentication. This option supplies the model client only; it does not inject credentials into functions or prompts.

Use an Adios agent locally

The Adios CLI can fetch a saved agent and run its workflow with a local model:

adios agent run AGENT_ID --local --model YOUR_MODEL \
  --model-url http://localhost:1234/v1 --input '{"name":"Adios"}'

This mode requires Adios login. Connector operations can use the agent's Adios connections, and saved-agent approvals remain in Adios. Using a remote connection does not require copying its raw credential into the kit.

For edits in a local directory, use --local --dir ./hello-workflow. Review the CLI's effect simulation behavior before using a local directory with write operations. Standalone recordings, local drafts, saved-agent runs, and hosted test runs have different effect policies; a successful simulation does not prove a real write.

Credentials and functions

The intended credential choices are explicitly configured local secrets and authorized credentials resolved from Adios. Today, local model keys and managed connector access are available through the paths above. General secret export or automatic injection into arbitrary functions is not yet part of this interface.

Function runners sanitize their environment. An ambient shell variable is not automatically a function secret. Keep credentials out of workflow inputs, recordings, artifacts, and trace files. Use a host-owned connection for service access.

Resume and inspect

Keep the state directory to resume a standalone run with the same run ID. Approval and question decisions are retained with that run. Use --approve later to leave approvals waiting, or --test for simulated effects with approvals and delays skipped. --trace records model calls and replies for debugging; those files can contain the business data supplied to the agent.

See functions and testing for native tests and Adios Sandbox's role.