Skip to content
agentFast
Reference

CLI

new, serve, run, eval, approvals, mcp.

Installed as agentfast with the package. Every command reads agentfast.yaml from the working directory unless --config says otherwise.

agentfast new

Scaffold a project from a template.

agentfast new support --sdk langgraph
agentfast new research --sdk vanilla --name my-researcher

| Option | Default | | |---|---|---| | template | — | support, research, or code | | --sdk | langgraph | Which orchestration to scaffold against | | --name | <template>-agent | Project directory name |

Available combinations: support on all five SDKs; research and code on vanilla.

agentfast serve

Run the API that the playground, dashboard and HITL queue all consume.

agentfast serve                      # reads agentfast.yaml
agentfast serve --port 8321
agentfast serve --in-memory          # no Postgres — demos only

| Option | Default | | |---|---|---| | --config, -c | ./agentfast.yaml | Path to config | | --host | 0.0.0.0 | | | --port | 8321 | | | --in-memory | off | Ephemeral stores. A restart loses every run |

agentfast run

Run the agent once from the terminal and print the outcome — no server, no browser.

agentfast run "I need a refund for ord_1001"

Useful in scripts and for smoke-testing a config change. A run that pauses for approval prints the approval id and exits.

agentfast eval

Run the quality gate. Exits non-zero below threshold.

agentfast eval
agentfast eval --case refund_damaged_item

See Evals.

agentfast approvals

Drive the HITL queue without a UI.

agentfast approvals list
agentfast approvals list --status pending

agentfast approvals approve appr_run_abc_call_1 --by alice@ops
agentfast approvals reject  appr_run_abc_call_1 --by alice@ops --note "already refunded"

Approving a suspended run resumes it. The rejection note goes back into the agent's reasoning — see Human-in-the-loop.

agentfast mcp

Manage MCP servers your agent consumes, and publish your agent as one.

agentfast mcp list
agentfast mcp add github --command npx --args "-y,@modelcontextprotocol/server-github"
agentfast mcp test github          # connects for real and lists its tools
agentfast mcp remove github

agentfast mcp skills               # playbooks exposed to MCP clients
agentfast mcp serve                # publish this agent as an MCP server

add and remove edit agentfast.yaml in place — the same operations the dashboard's MCP panel performs. See Tools and MCP.

TipEverything is also HTTP

Every CLI operation has an API equivalent, because the CLI, the dashboard and the playground all drive the same runner. If you're automating, the HTTP API is usually the better surface.