Deleting a workflow hands back a short-lived confirm token, edits to a live workflow land in a draft rather than the pipeline, and provisioning a use-case kit expects its preview to have been reviewed first. The step-level tools are the recommended path for a reason the descriptions state outright — adding or updating a step returns per-step validation errors immediately, while pushing a whole steps array silently strips root fields. Cost is made visible rather than implied: the credit tools label the exact window every total covers, each model carries its own credit price, and the code test is marked as consuming none.
The workspace-wide control surface for Agentled: workflow authoring and publishing, executions down to the individual step, agents and their scheduled routines, knowledge lists and the graph over them, and the use-case records that tie the rest together. One key and one credit balance stand in for the dozens of service subscriptions a workflow would otherwise need — enrichment, email finding, scraping, models, image and video generation — each with its own credit price.
- Authoring one step at a time: create_workflow, then add_step, update_step, move_step and remove_step, each returning that step's validation errors as you go, with validate_workflow reporting broken next-step references, missing required fields, unreachable steps and invalid app or action IDs before publish_workflow takes it live.
- A draft between your edits and the live pipeline: changes to a live workflow land in a draft snapshot, get_draft inspects it, promote_draft overwrites the live config with it, and discard_draft throws it away.
- Config history you can walk back: list_snapshots pages through automatic and manual snapshots, get_snapshot_content reads one without restoring it, restore_snapshot reverts steps, context, name, description, goal and style, and create_snapshot saves a checkpoint before a risky change.
- Execution reading at step granularity: start_workflow returns an execution id, get_execution maps each step to its output, list_timelines and get_timeline give each step's record with its computed prompt, and read_step_output pulls an output a prompt deferred as a tool reference.
- Reruns without rebuilding: rerun takes only a timeline id and derives the workflow, execution and step from it.
- Isolated testing before anything is wired in: test_app_action runs one app action against input data, test_ai_action runs a prompt template with variables and a response structure, and test_code_action executes JavaScript in the same sandboxed context the production orchestrator uses.
- Knowledge lists as typed tables: create_knowledge_list defines the schema, upsert_knowledge_rows writes in bulk resolving by id or user key, get_knowledge_rows pages through them, and snapshot_knowledge_list with restore_knowledge_list_snapshot gives a self-contained backup and a merge-restore that keeps computed fields.
- A graph over that data: query_kg_edges traverses relationships and returns scored source and target nodes, and get_scoring_history returns past decisions with their scores and dates — so a scoring workflow compounds on previous runs rather than starting cold.
- Agents with files, skills and schedules: create_agent and update_agent set instructions, enabled apps and actions, model tier and a daily credit ceiling; upload_agent_file attaches content; create_routine schedules autonomous runs on named intervals, bounded by steps and credits per run.
- Inbound channels: list_channels reports email, Slack, WhatsApp, Signal and Telegram configuration, configure_channel sets allowed senders and outbound policy, and set_channel_defaults caps sessions per day and session length.
- Credit accounting with its window attached: get_workspace_credits and get_workflow_credits each label the exact period a total covers, with cost drivers available on request.
- Ways in that are not pipeline JSON: chat builds a workflow through conversation, and do routes a plain-English intent to the best-matching live workflow with a confidence score.
- Import and transfer: preview_n8n_import returns the mapped step graph, unsupported nodes and remediation without creating anything, import_n8n_workflow lands the result as a draft, and export_workflow with import_workflow moves a workflow between workspaces.
- Work handed to an external builder: create_builder_work_item makes a durable handoff, claim and submit move it along, and review sets accepted, needs-changes or closed.
- Output shared outside the workspace: create_public_form_link opens a workflow's input form to people without an account, with an expiry and a submission limit, and share_execution returns a share URL for chosen outputs of one run.
An Agentled workspace and its API key in AGENTLED_API_KEY, generated under Workspace Settings. Two install paths exist — registering the server directly, or the Claude Code plugin that bundles the server with the workflow-authoring skill — and you should pick one: doing both leaves two identical server processes and the skill registered twice. Credits are the running cost, and scope decides what a key may do: some administrative tools are refused unless the caller holds the matching scope, and routines are a paid feature a free workspace cannot create or trigger.
One command plus a key — claude mcp add --transport stdio --scope user agentled \ -e AGENTLED_API_KEY=wsk_... \ -- npx -y @agentled/mcp-server, then supply credentials