Vercel Adds ACP-compatible harnesses to Its AI SDK Harness Layer.

Vercel is making its AI SDK harness layer less dependent on one-off runtime integrations. In an August 13 changelog update, the company said developers can now connect any Agent Client Protocol-compatible harness to HarnessAgent using a new package, @ai-sdk/harness-acp.

The change introduces a protocol-level adapter rather than another adapter built around a single coding agent. That distinction matters for teams assembling agent systems across different runtimes. If a harness ships an ACP-compatible package, developers can map it into Vercel’s harness layer instead of waiting for Vercel to create and maintain a dedicated integration.

Until now, every AI SDK harness adapter wrapped a particular runtime. Vercel lists Claude Code, Codex, Pi, Deep Agents and OpenCode among the existing examples. The new @ai-sdk/harness-acp package wraps the protocol itself.

Vercel describes this as a meta adapter. Developers pass an ACP-compatible package to the createACP function, then provide the basic mapping needed for the harness. The resulting adapter can be used through HarnessAgent, keeping the higher-level AI SDK interface consistent while changing the runtime underneath it.

The company’s example configures a Codex ACP implementation with:

• A unique harness identifier, such as codex-acp
• An npm package source, in this case @agentclientprotocol/codex-acp
• The executable that launches the harness
• Environment variables that should be forwarded
• A mapping between AI SDK permission modes and runtime-specific session modes

That permission mapping is an important part of the integration. In Vercel’s sample, read and edit modes do not require a special ACP session mode, while allow-all maps to a runtime mode called agent-full-access. The example also forwards CODEX_API_KEY and OPENAI_API_KEY, showing how credentials can be made available to the underlying executable without changing the surrounding agent interface.

import { createACP } from '@ai-sdk/harness-acp';

export function createCodexACP() {
  return createACP({
    harnessId: 'codex-acp',
    source: {
      type: 'npm-simple',
      packageName: '@agentclientprotocol/codex-acp',
    },
    executable: 'codex-acp',
    forwardEnv: ['CODEX_API_KEY', 'OPENAI_API_KEY'],
    permissionModeMapping: {
      'allow-reads': null,
      'allow-edits': null,
      'allow-all': {
        type: 'session-mode',
        modeId: 'agent-full-access',
      },
    },
  });
}

Agent tooling is still fragmented. Coding agents often expose different launch commands, authentication requirements and permission controls, even when they perform similar tasks. Dedicated adapters can smooth over those differences, but each new runtime creates another integration that must be written, tested and updated.

ACP support shifts part of that work toward a shared interface. For small product teams, the immediate benefit is optionality: an ACP-compatible runtime can potentially be introduced without redesigning the application around its proprietary API. It also gives harness authors a clearer route into the AI SDK ecosystem, provided they publish a compatible package and define the necessary mappings.

The update does not eliminate integration work. Developers still need to specify the executable, package source, environment forwarding and permission behavior. Compatibility also depends on the selected harness actually shipping an ACP-compatible package. The improvement is that these details can now be expressed through one reusable adapter model rather than a new AI SDK adapter architecture for every runtime.

This release extends Vercel’s broader effort to make the harness layer support more coding agents. It follows the company’s addition of Grok Build to the AI SDK harness layer, but takes a more general approach by targeting a protocol instead of one provider.

What is Vercel ACP-compatible harnesses?

Vercel ACP-compatible harnesses are integrations that connect Agent Client Protocol-compatible agent runtimes to the AI SDK’s HarnessAgent. They are created with the @ai-sdk/harness-acp package and the createACP function, using configuration that maps the runtime’s package, executable, environment and permissions.

Who should use the new ACP harness package?

It is intended for developers and harness authors who want to connect an ACP-compatible runtime to Vercel’s AI SDK harness layer. It is particularly relevant when no dedicated AI SDK adapter exists for the selected runtime.

When are ACP-compatible harnesses available?

Support is available now through the AI SDK’s @ai-sdk/harness-acp package, according to Vercel’s August 13, 2026 changelog. Vercel did not specify separate pricing or plan requirements in the announcement.