Agent Client Protocol

Definition

Agent Client Protocol (ACP) is presented in the DeepChat source as a protocol boundary through which a client can run or interact with external agents and their harnesses without embedding every harness implementation directly in the product.

Current Synthesis

ACP addresses agent-to-client interoperability, while MCP addresses access to tools, data, and context. In DeepChat’s architecture, this distinction lets the client keep one native Tape and harness while still offering external agents through a common integration surface. Protocol-level integration can reduce direct coupling and make the client a shared human-interaction and orchestration layer.

The evidence remains early and implementation-specific. The maintainers report missing functionality and incompatible vendor extensions, so ACP should be treated as an emerging interoperability strategy rather than a settled universal standard.

Key Claims

• A client can support external agents without internally maintaining several complete harness stacks.
• Protocol boundaries can reduce coupling between a user interface and rapidly changing agent runtimes.
• ACP and MCP solve different layers: external-agent integration versus tool and context connectivity.
• Shared client surfaces can support orchestration, human interaction, and virtual-worker products across multiple agents.
• Interoperability weakens when implementations add incompatible extensions or omit required capabilities.

Evidence

DeepChat integration choice

• VOL.002｜DeepChat：为什么要做一块开源 AI 试验田？ says DeepChat intends to retain its own Tape and harness while using ACP to run external agents.

Adoption and immaturity

• VOL.002｜DeepChat：为什么要做一块开源 AI 试验田？ reports ACP use in multi-agent orchestration, human interaction, and virtual-employee products while also noting incomplete functions and divergent extensions.

Counterevidence & Qualifications

• The source does not define the full protocol, compare implementations, or demonstrate compatibility through independent tests.
• A protocol can standardize messages without standardizing permissions, memory, tool semantics, observability, or result quality.
• Vendor-specific extensions may recreate the coupling the protocol is intended to avoid.

What Changed

• Created a distinct interoperability concept for external-agent/client integration.
• Separated ACP’s role from MCP’s tool-access role.
• Added immaturity and extension divergence as explicit limits.

Related Concepts

• Model Context Protocol - exposes tools, data, and context rather than complete external agents.
• Agent Harness - agent runtime that ACP may connect to a client.
• Multi-Agent Collaboration - orchestration setting that can benefit from shared client integration.
• Human-Agent Collaboration - interaction layer a common client may provide across agents.
• DeepChat - source project using ACP as its external-agent boundary.