Creating agents

Creating agents requires the creator or admin role. Each person can create a limited number of agents by role (see Usage limits), and your plan caps the total.

The New agent dialog

Click Create Agent in the Dashboard header, or Use template in the Library. Pick a source on the left:

Source	What you get
Blank agent	An empty agent with default instructions. Add skills, credentials and schedules afterwards. You choose its runtime and optionally a model.
Import a GitHub repository	An agent built from a repository you already have (see below).
Starter templates	Curated starting points such as scout (research), sage (analysis) and scribe (writing).
Admin-configured templates	Templates your administrators curated for your company.

Then enter a Name, for example Market Scout. The dialog previews the permanent ID (market-scout) — lowercase with hyphens — used in URLs, API paths, MCP tool names, schedules and webhooks. The name you typed becomes the agent’s display label, which you can change later; the ID cannot change except through Rename.

Click Create agent. The container is created and started, and the dialog runs a compatibility check and shows its result.

On ExoHuman Cloud, a new agent runs on your instance’s default compute profile unless you assign another.

Importing a GitHub repository

When you import a repository, choose how ExoHuman takes it on:

Intent	What happens	Sync
Clone (default)	Clones the repository and keeps it connected to that remote.	Yes — on the agent’s own working branch when the repository is yours; pull-only otherwise.
Fork	Forks the repository into your GitHub account first (the template must support it).	Yes — to your fork.
Copy	Takes a one-time snapshot of the files without git history.	No — a standalone workspace.

For Clone, ExoHuman asks What is this repository?: an agent (the repository is the agent’s memory, skills and state, saved to its own branch) or a deployment of a codebase (a product the agent works on, which it only pulls). Private repositories need a GitHub token (Platform keys).

Template format

A template is a folder or repository with:

File	Purpose
template.yaml	Metadata: display name, description, runtime, resources, required credentials, schedules, use cases
AGENTS.md	The agent’s instructions (CLAUDE.md is accepted for compatibility)
.mcp.json.template	MCP server configuration with ${VAR} placeholders filled from credentials
.env.example	The environment variables the agent expects
.claude/skills/	Skills shipped with the agent

Schedules declared in template.yaml are created with the agent (disabled until you turn on autonomy).

display_name: Market Scout
description: Monitors competitors and writes a daily briefing.
runtime:
  type: pi
resources:
  memory: 2g
  cpu: 1
use_cases:
  - What did our competitors announce this week?
  - Compare our pricing with Competitor A.
schedules:
  - name: daily-briefing
    cron: "0 8 * * 1-5"
    timezone: Europe/Madrid
    message: Write today's competitor briefing and publish it as a report.

Creating agents by API or MCP

POST /api/agents
Authorization: Bearer <token>
Idempotency-Key: 3f1c…
Content-Type: application/json

{ "name": "market-scout", "display_label": "Market Scout", "template": "github:acme/market-scout", "import_intent": "clone" }

The Idempotency-Key header makes retries safe. MCP clients can use the create_agent tool. See Agents API.

Managing agents

From the agent page header and the Dashboard you can:

• Start and Stop the container;
• set the display label, avatar and tags;
• Rename the agent (changes its ID and updates references);
• Delete it (owner or admin) — its container, workspace and schedules are removed.

The Overview tab shows health, recent activity, cost and context usage. Agent page tabs are: Overview, Tasks, Chat, Dashboard (when the agent defines metrics), Reports, Canvas, Schedules, Loops, Skills, Sharing, Permissions, A2A (when enabled), Git (when synced), Files and Settings.

Ephemeral agents

Orchestrating agents can spawn short-lived ephemeral helpers for a single job and discard them afterwards. Ephemeral agents are an instance module.