Call Summary Agent with Sandbox
https://github.com/vercel-labs/call-summary-agent-with-sandbox- Category
- AI Agents
- Rank
- No. 1448Tools index
Previous survey · No. 1454 ·
- Pricing
- Open Source
- Type
- AGENT
- Builder
- vercel-labs
- GitHub
- 58 stars
- Date
About
Demo agent that summarizes calls and runs follow-up code in Vercel Sandbox.
What it does
It turns Gong call transcripts into sales analysis. A webhook starts a durable workflow, loads transcript and account context into an isolated environment, and lets an AI agent search those files with shell commands before writing a summary, owned tasks, objections, and handling assessments.
Why it's ranked here
This is a credible reference implementation, not a finished sales product. It combines a working webhook path, transcript retrieval, isolated file exploration, streamed progress, mock data, and retryable workflow steps. Its narrow Gong starting point and text-based result keep the production burden visible.
What's good
Demo mode runs without Gong credentials and includes a substantial sample transcript, prior calls, research, competitive context, CRM records, and a sales playbook. Production mode fetches Gong transcripts, maps speakers, preserves timestamps, and logs sandbox commands and truncated output for inspection.
Tradeoffs
The whole agent run sits inside one workflow step, and the source itself recommends finer durable checkpoints for high-volume production. The implementation returns model text rather than enforcing the documented structured schema. Missing Gong credentials silently select demo data, which can hide configuration mistakes.
How to use it well
Use it as a starter for engineers prototyping post-call sales analysis on Vercel, especially when file-based context and observable agent exploration matter. Replace the prompt, context set, model, and transcript adapter for your workflow. It does not provide built-in CRM updates, Slack delivery, or connectors beyond Gong.
Technical notes+
app/api/gong-webhook/ is the documented POST entry point, while next.config.ts wraps Next.js with Workflow support. lib/agent.ts configures ToolLoopAgent, creates a Vercel Sandbox with a ten-minute timeout, streams log entries, and returns result.text. lib/tools.ts uploads generated files through bash-tool and records commands, output previews, and nonzero exits. lib/sandbox-context.ts converts transcript data into Markdown plus metadata and adds demo context. lib/gong-client.ts calls Gong's transcript endpoint with Basic authentication. lib/config.ts defaults to anthropic/claude-haiku-4-5 and activates demo mode whenever either Gong credential is absent. package.json marks the Next.js application private and provides development, build, start, and lint scripts.
Observed
- Primary language
- TypeScript, with TS and TSX application sources
- Packaging
- Private Next.js application installed through pnpm; package metadata also supplies npm-compatible scripts
- Interface
- HTTP POST webhook endpoint for Gong payloads, with an empty JSON body supported in demo mode
- Deployment surface
- Vercel one-click deployment and local Next.js development are documented
- External integration
- Built-in Gong transcript retrieval using access-key and secret-key credentials
- Execution model
- Vercel Sandbox hosts file exploration through bash-tool inside a Workflow SDK process
- Testing structure
- The provided project tree lists no test directory or test script
Read from README.md, package.json, lib/agent.ts, lib/tools.ts, lib/types.ts, lib/config.ts, lib/logger.ts, lib/mock-data.ts, lib/gong-client.ts, lib/sandbox-context.ts, next.config.ts, postcss.config.js, app/page.tsx, app/layout.tsx, app/HomePage.tsx.
What it can do
Summarize call recordings
Call audio or transcript → Text summary of call content
Execute code in sandbox environment
Code snippets or scripts → Code execution results
Generate follow-up actions from call content
Call summary or transcript → List of actionable next steps
Run code demonstrations
Demo code requirements → Live code execution in sandbox
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.