Devin SDK example: harness chat

Source: harness-chat

The AI SDK harness example. Devin edits files, runs commands and calls your server-side tools in an isolated microVM; the chat shows each step as it streams, asks you to approve shell commands, and keeps its session so a later message — from any server instance — picks up where it left off. Side panels show the sandbox status, a file browser, and a preview of whatever Devin serves from the workspace.

Harness chat: a chat column with Devin's reasoning, edits and approved bash commands; side panels showing the Vercel Sandbox status, a preview of the built app, and its files — The console: chat with Devin's reasoning and tool calls, the sandbox, a live preview of the app it built, and its files

How it fits together

The whole agent is one declaration:

lib/agent.ts (excerpt)

import { HarnessAgent } from "@ai-sdk/harness/agent";
import { createVercelSandbox } from "@ai-sdk/sandbox-vercel";
import { createDevin } from "@cognition-ai/harness-devin";

export const devinAgent = new HarnessAgent({
  harness: createDevin({ port: BRIDGE_PORT }),
  sandbox: createVercelSandbox({ runtime: "node24", ports: [BRIDGE_PORT, ...PREVIEW_PORTS] }),
  instructions: "You are Devin working in an empty workspace inside a sandbox. …",
  permissionMode: "allow-edits", // edits go through; shell commands ask, as a card in the chat
});

File	Role
`lib/agent.ts`	The agent above. Add AI SDK `tools` here and they run on your server, relayed into the sandbox.
`app/api/chat/route.ts`	The `useChat` route. Resumes the chat’s session, streams the turn, parks the session (`detach()`) when the stream settles.
`lib/session-store.ts`	Resume state in Redis, keyed by chat id, which is also the harness `sessionId` and therefore the sandbox name.
`lib/workspace.ts`	Read-only view into a chat’s sandbox (status, files, preview URL) for the side panels.

Run locally

Local development uses the real Vercel Sandbox; there is no emulation, so what you see locally is what runs in production.

git clone https://github.com/COG-GTM/devin-sdk-examples
cd devin-sdk-examples/harness-chat
bun install                    # @cognition-ai/sdk + harness-devin from npm (the pinned beta)
vercel link                    # pick or create the project (once)
vercel env pull                # .env.local: VERCEL_OIDC_TOKEN, KV_REST_API_*, REDIS_URL
echo 'DEVIN_API_KEY=cog_…' >> .env.local
bun run dev                    # http://localhost:3210

vercel env pull needs the Upstash integration on the project (Storage → Upstash in the Vercel dashboard). The OIDC token expires; rerun vercel env pull when sandbox calls start failing with an auth error.

Deploy

The README has a Deploy-with-Vercel button that clones the example, provisions the Upstash store and asks for DEVIN_API_KEY; Vercel Sandbox authenticates automatically on Vercel. @cognition-ai/harness-devin and @cognition-ai/sdk are pinned to the same published beta in its package.json.

This is a single-user demo: there is no login, and anyone who can reach the deployment can open any session in it. Put it behind Vercel Deployment Protection or add your own auth before sharing a URL.

AI SDK harness

What @cognition-ai/harness-devin provides and how the bridge works.

Diagnostics

The onLog events this example turns into a status line.

Get started

Sessions

Integrations

Examples

SDK reference

API client

Resources

Devin SDK example: harness chat

How it fits together

Run locally

Deploy

AI SDK harness

Diagnostics

​How it fits together

​Run locally

​Deploy

AI SDK harness

Diagnostics

How it fits together

Run locally

Deploy