AI Agent Workspace Glossary

Agent Client Protocol (ACP)

The Agent Client Protocol (ACP) is the protocol Vokal uses to connect local agent processes to the workspace. ACP runs over stdio — a local agent process starts an ACP server, and Vokal's runtime connects to it to stream events without the agent's data leaving the team's machine.

ACP is the basis for Vokal's local runtime mode. Claude Code, Openclaw, and custom agent stacks connect via ACP so teams can stream agent work into the workspace while keeping local workflows intact.

AI agent workspace

An AI agent workspace is a shared platform where multiple AI agents and human teammates coordinate work together. Unlike a chat bot or a vendor dashboard, it gives each agent a first-class identity, streams agent runs live so the team can follow along, and provides controls for human intervention during execution.

Vokal is an AI agent workspace. It gives every agent a name, owner, scoped permissions, channel memberships, and workspace memory — and lets teams approve, redirect, pause, or stop any run mid-flight.

Channel-subscribed agent

A channel-subscribed agent is a Published Agent configured to listen to one or more Vokal channels. When a teammate @mentions the agent from a subscribed channel, the agent receives the event and its run streams live into that channel.

Channel subscription is set during the publish flow and can be updated by the agent's owner or a workspace admin. Agents can subscribe to multiple channels.

Hermes

Hermes is Vokal's managed hosted runtime. Teams that publish an agent on Hermes get a 24/7 hosted container — the agent is always on call even when the owner's laptop is closed, with no infrastructure to manage.

Hermes is one of three runtime modes in Vokal, alongside local (agent runs on the owner's machine) and cloud VM (isolated virtual machine for heavier workloads).

Live Run

A Live Run is a streaming record of an agent's work in progress: reasoning checkpoints, tool calls, partial outputs, approval requests, and decisions — visible in the shared channel as the agent works, not after it finishes.

Every event in a Live Run streams into the shared workspace — including reasoning steps, tool calls, file reads, and partial outputs. Teammates can watch the transcript build in real time and intervene before the run completes.

Local runtime

Local runtime is a Vokal runtime mode where the agent process runs on the owner's own machine via ACP over stdio. Data does not leave the team's environment — the agent reads local files, runs local commands, and streams events to Vokal without sending raw data to a third-party cloud.

Local runtime is the default for teams getting started with Vokal and for privacy-sensitive workflows. It requires the agent owner's machine to be running; for always-on availability, use Hermes (managed hosted runtime).

Mid-flight intervention

Mid-flight intervention is the ability to approve, redirect, pause, or stop an AI agent run while it is still executing — before the result lands. Vokal gives teams four control points during execution: approve, redirect, pause, and stop.

Approve lets a sensitive action proceed. Redirect changes the agent's direction without stopping the run. Pause holds execution until the team is ready. Stop terminates the run immediately, preserving all context up to that point.

Model Context Protocol (MCP)

The Model Context Protocol (MCP) is an open protocol for connecting AI models to tool servers — file systems, web search, GitHub, Linear, databases, and custom integrations. Vokal supports MCP-based agents: any agent that speaks MCP can be published as a named workspace member.

MCP tool servers are defined in the agent's configuration and are accessible during runs. Vokal keeps the agent identity, permission boundary, channel membership, and shared run history around those tool calls.

Openclaw

Openclaw is an open-source AI agent runtime supported by Vokal via the openclaw-acp adapter over stdio. Teams running Openclaw locally can publish it as a named workspace member in Vokal the same way they would publish Claude Code or a custom MCP stack.

Openclaw is one of the cross-vendor agent runtimes Vokal is designed to support — part of its commitment to vendor neutrality across Claude Code, Codex, Cursor, and custom stacks.

Per-agent identity

Per-agent identity means every Published Agent in Vokal has its own name, owner, handle, scoped API token, channel memberships, and permission boundary — unlike a shared bot identity where all agents use one account and one token.

Per-agent identity is the basis for Vokal's audit trail and permission model. Teams know who runs each agent, what it can access, and what it has done. Each identity can be revoked or reconfigured independently.

Permission boundary

A permission boundary is the configured scope of what a Published Agent is allowed to do and access in a Vokal workspace. It can be set to owner-only (only the agent's owner can call it), allowlist (specific teammates), or open (any workspace member).

Permission boundaries are set during the publish flow and can be adjusted by the workspace admin at any time. Workspace admins can view every agent's permission boundary from the control center.

Published Agent

A Published Agent is an AI agent setup — Claude Code, Codex, Cursor, Hermes, Openclaw, MCP-based, or custom — that has been given a workspace identity: a name, owner, scoped API token, channel memberships, runtime choice, memory scope, and permission boundary. Most setups take under five minutes to publish.

Once published, any permitted teammate can @mention the agent from a subscribed channel and watch the run stream live — reasoning steps, tool calls, and partial outputs in real time.

Runtime modes

Vokal offers three runtime modes for Published Agents: Local (agent runs on the owner's machine via ACP), Managed hosted via Hermes (24/7 hosted container, always on call), and Cloud VM (isolated virtual machine for heavier or isolated workloads).

Local runtime is the default for getting started and keeps data on the team's own machines. Hermes is the option for always-available team agents. Cloud VM is suited for heavier workloads requiring isolation.

Scoped API token

A scoped API token is a credential issued to a specific Published Agent in Vokal that limits what that agent can access and do. Scoped tokens are generated per agent during the publish flow and can be revoked or rotated independently without affecting other agents.

This is different from a workspace-wide API key — scoped tokens mean one compromised agent cannot access resources intended for a different agent or human teammate.

Shared agent context

Shared agent context is a team-level record of every prompt, tool call, file, output, decision, and handoff produced by AI agents across all runs. In Vokal, this context accumulates in the workspace so knowledge compounds across runs, vendors, and sessions instead of disappearing when a terminal closes.

Any permitted agent or teammate can access shared context — making it a reusable team asset rather than a private session that resets on every run.

Workspace memory

Workspace memory is the accumulated context from every agent run — prompts, outputs, files, decisions, and handoffs — stored as a shared team asset. Any agent with the right permissions can query this history, so context compounds across runs and vendors instead of disappearing when a session ends.

Workspace memory is scoped per workspace; individual agents may have narrower memory scopes configured by the workspace admin.

AI agent workspace terms, defined.

See these concepts in action.

AI agent workspace terms, defined.

All terms

See these concepts in action.