Anter MCP Server

Agent lifecycle tools

These tools cover the design loop: learning the spec format, creating a draft, revising it, and reading it back — the same lifecycle described in Flows key concepts, driven through MCP instead of the dashboard.

Discovery tools

Call these first — they cost nothing but a tool call and don't touch any of your data.

ToolWhat it returns
get_agent_lifecycle_guideA cross-tool orchestration guide: which tool to call in what order, from prerequisites through promotion
get_spec_schemaThe AgentSpec JSON Schema, plus a prose breakdown of the model (agent → sub-agents → connections → skills → tools)

Start here

If your MCP client hasn't designed an Anter agent before, have it call get_agent_lifecycle_guide before anything else — it's a static reference that lays out the whole sequence in one call.

Creating and revising

create_draft_agent{ projectId, spec, idempotencyKey }

Persists a new agent as a draft: a version 1 revision, not yet promoted to live. Nothing about this call goes to production traffic. Re-submitting the same idempotencyKey returns the existing draft instead of creating a duplicate — safe to retry.

update_agent_spec{ agentId, projectId, spec, idempotencyKey }

Appends a new revision to an agent's history, whether it's still a draft or already live. This never touches what's currently serving live traffic — it prepares the next version for review and testing. Like create_draft_agent, it's idempotency-key safe to retry.

Promotion is dashboard-only

There is no MCP tool that makes a draft or revision live. Once a spec looks right, promote it from the Anter dashboard (VersionsPromote). This is a deliberate boundary — it keeps "what's live" a conscious, UI-driven decision, not something a scripted MCP call can trigger silently. See Versions for the promotion model.

Reading agents back

get_agent_by_slug{ agentSlug }

Looks up an agent by its human-readable slug (the same one you set when creating it) instead of its internal id. Returns the agent's current spec — whichever revision is "current" (a pending draft revision if one exists, otherwise the active version) — ready to feed straight into update_agent_spec. A slug with no match returns { found: false } rather than an error.

get_agent_structure{ agentId }

Returns the full agent graph: the root agent, every sub-agent, and how tools and skills are wired to each — the same structure you'd see on the agent graph canvas in Flows.

A typical loop

1

Learn the format

Call get_agent_lifecycle_guide and get_spec_schema once to understand the AgentSpec shape and the recommended tool sequence.

2

Draft the agent

Assemble a spec and call create_draft_agent. You get back an agentId you'll reuse for every following call.

3

Test it

Use run_agent against the draft to see how it behaves before anything is live.

4

Iterate

Call update_agent_spec with the same agentId to append a revised version. Repeat as many times as needed.

5

Promote

When you're satisfied, promote the version from the dashboard. This step has no MCP equivalent by design.