Concepts

Terms, file layout, and how the pieces fit together

The agent platform has a handful of moving parts, and they don’t always have obvious names. This page is your reference: every term, every standard file path, every reserved port. Skim it now to get the lay of the land; come back later when something in the docs reads like jargon.

The shape of things

Workspace (your team)  ──►  Agent (container)  ──►  Channels / Routes / Tasks
   │                            │
   │                            ├─ Skills (capabilities)
   │                            ├─ Secrets (env vars)
   │                            ├─ Models (LLMs it can call)
   │                            └─ Workspace files + snapshots
   ▼
Pinata IPFS + R2 (storage)

A workspace is a Pinata team. Inside it you create one or more agents — each is an isolated container running an agent engine, with its own files, gateway, and per-agent subdomain.

Glossary

Agent

A single isolated container with a persistent workspace, gateway, and {agentId}.agents.pinata.cloud subdomain. Created via POST /v0/agents or the Create Agent button. Identified by agentId (a short slug like x0i33jye).

Engine

The runtime that powers the agent. Engines are enabled per deployment:

Engine	Tagline	When to pick it
openclaw (default)	Configurable agent for custom workflows and power users	You want full control over manifest fields, scripts, routes, and skills
hermes	Opinionated agent with a smoother out-of-the-box experience	You want fewer knobs and first-run channel setup to feel more direct
super-builder (beta)	Headless coding agent optimized for issue-board work	You’re assigning work through Issues & Boards and want longer runs (200-turn default cap) over chat features. Chat-first surfaces like Channels aren’t available.

Pick the engine when creating the agent — either in the Agent Engine card on the Create Agent wizard, or by passing engine in POST /v0/agents, or by setting engine in manifest.json for a template. Available engines per deployment are listed at GET /v0/agents/engines.

Gateway

The long-running process inside each agent container that:

• Terminates WebSocket connections from the dashboard, channels, and CLI
• Routes path/domain traffic to user-defined ports
• Brokers HTTP API calls for the per-agent subdomain
• Restarts on Restart Gateway in the Danger tab

Listens on the reserved port 18789.

Gateway token

Per-agent credential used to authenticate against the agent’s own subdomain ({agentId}.agents.pinata.cloud/...). See API → Gateway token. Distinct from the Pinata JWT (workspace-wide API key) and from the platform JWT (1h token an agent gets from itself via POST /v0/platform/token).

Token	Where it works	What it’s for
Pinata JWT	agents.pinata.cloud	Workspace-wide management - create/list/delete agents, manage secrets, browse templates
Gateway token	{agentId}.agents.pinata.cloud	Talk to one specific agent (chat, files, routes, console, git)
Platform JWT	agents.pinata.cloud	A short-lived JWT the agent gets for itself - lets the agent self-modify via @pinata/platform

Workspace

Two meanings - context disambiguates:

1. Team workspace - your Pinata account or shared team. Switch under Account → Workspaces.
2. Agent workspace - the per-agent file tree inside the container. OpenClaw agents use /home/node/clawd/workspace/; Hermes agents use /home/hermes/data/workspace/. The workspace includes manifest.json, SOUL.md, attached skills under skills/, file uploads under uploads/, and anything else your agent has written.

Workspace anatomy

The default file layout inside workspace/:

File	Purpose
manifest.json	Agent configuration - identity, secrets, skills, scripts, routes, channels, tasks. Single source of truth.
SOUL.md	Agent personality and principles - customize freely
AGENTS.md	Conventions the agent follows when working in the repo (memory system, safety rules, file layout)
IDENTITY.md	Name, vibe, emoji, owner identity - written at deploy time
USER.md	Notes about the human user, learned over time
TOOLS.md	Environment notes (what tools exist in the container)
BOOTSTRAP.md	First-run conversation guide. Self-deletes after setup.
HEARTBEAT.md	Periodic-task notes (empty by default)
skills/<name>/	Files for each attached skill
uploads/	Files uploaded via the chat attachment button or POST /v0/agents/{agentId}/files/upload

You can commit and push anything else (projects/, scripts/, etc.) - the agent treats the whole tree as fair game.

manifest.json vs runtime config

These are different files. They live in different places and have different schemas.

File	Path	Schema	Edited by
manifest.json	<workspace>/manifest.json	manifest.v1.json	You (committed to git, validated via POST /v0/templates/validate)
OpenClaw runtime config	/home/node/.openclaw/openclaw.json	OpenClaw runtime schema	OpenClaw at runtime (validated via POST /v0/agents/{agentId}/config/validate)
Hermes runtime config	/home/hermes/data/config.yaml	Hermes runtime schema	Hermes at runtime (edited from the Danger tab’s Config row)

Day-to-day, edit manifest.json. Runtime config is the engine’s own view of that configuration and is usually only touched from the Danger tab when you’re debugging the engine itself.

Skill

A reusable package of files and instructions that extends an agent’s capabilities. Pinned to IPFS, addressed by slug (@pinata/api) or CID. Installed into your Skills Library, then attached per-agent. Up to 10 attached per agent. See Skills.

The platform ships with @pinata/platform bundled - it lets the agent perform self-service operations (install skills, set secrets, manage tasks).

Secret vs variable

Both are environment variables injected at container start. The difference:

Kind	Storage	Visible in API?	Use for
Secret	AES-GCM encrypted with a per-user key	Never returned	API keys, tokens, passwords
Variable	Plaintext	Returned in listings	Public URLs, feature flags

See Secrets.

Channel

A messaging surface the agent can be reached on - Telegram, Slack, Discord (WhatsApp coming). Configured per-agent. See Channels.

Route

A mapping from an external URL prefix or domain to a container port. Lets you expose web apps and APIs running inside the agent. See Routes & Domains.

Task

A scheduled prompt or system event. Three schedule kinds: at (one-shot), every (interval), cron. See Tasks.

Lifecycle script

Shell commands run by the agent runner at well-defined points:

Script	When	Working dir	Log file	Timeout
build	After deploy, after each git push, on Retry Scripts	/home/node/clawd	/tmp/user-build.log	5 min
start	After successful build, on agent boot	/home/node/clawd	/tmp/user-start.log	None (detached)

Defined in scripts in manifest.json.

Snapshot

A capture of the agent’s workspace pinned to IPFS as a single CID. Created automatically every minute when changes are detected, plus on demand via POST /v0/agents/{agentId}/snapshots/sync. Reset to any historical snapshot via POST /v0/agents/{agentId}/snapshots/reset. See Files & Snapshots.

Device

A client paired with an agent (CLI, mobile, browser session) - approval-gated for security. Listed at GET /v0/agents/{agentId}/devices. See Danger → Devices.

Template

A pre-configured agent (manifest + workspace files + skill list) packaged for one-click deployment. Published on the Marketplace or kept private in My Templates. See Templates.

Issue

Kanban-style work item assigned to an agent - title + prompt + optional repo. The agent runs it in an isolated <workspace>/runs/<runId>/ directory (its own git history; ignored by snapshots), you review the diff, then accept, revise, or open a PR. Boards carry per-board instructions, model overrides, and turn/cost caps. See Issues & Boards.

Filesystem reference

Paths agents and developers reach for most often inside the container:

Path	What’s there
/home/node/clawd/	OpenClaw runner root - where build/start execute
/home/node/clawd/workspace/	OpenClaw workspace
/home/hermes/data/workspace/	Hermes workspace
<workspace>/skills/<name>/	Attached skill files
<workspace>/uploads/	Uploaded files
<workspace>/runs/<runId>/	Isolated working dir for each issue run (gitignored, excluded from snapshots)
/home/node/.openclaw/openclaw.json	OpenClaw runtime config
/home/hermes/data/config.yaml	Hermes runtime config
/tmp/openclaw/openclaw-YYYY-MM-DD.log	Daily OpenClaw log file
/tmp/user-build.log	build script output
/tmp/user-start.log	start script output

Reserved ports & names

Reserved	Why
Port 18789	Gateway listens here
Subdomains containing pinata	Reserved for first-party use
Specific reserved domain suffixes	Reserved for the platform

Path routes must use ports between 1025 and 65535, excluding 18789. See Routes.

Identifier formats

Quick reference for the IDs you’ll see in URLs and API responses:

Identifier	Format	Example
Agent ID	Random short slug	x0i33jye
Snapshot CID	IPFS CIDv1, bafy... or Qm...	QmUMfo19uXMdBSXLiZAz7w...
Skill CID	IPFS CIDv1, bafy...	bafybeicglyjdb6w...
Skill slug	@<owner>/<name> (lowercase, hyphens)	@pinata/api
Custom domain ID	UUID	2fcd2a0b-aa70-432d-a310-678e01570e65
Secret ID	Random opaque string	(server-assigned)

Where to go next

Manifest reference

Every field in manifest.json

HTTP API

Auth, endpoints, and examples

Troubleshooting

Debug a stuck agent

Error reference

Look up an API error