Tabstack TypeScript Integrations

Long-term home for every Tabstack TypeScript integration. Each adapter wraps the official Tabstack TypeScript SDK (@tabstack/sdk) and exposes Tabstack's extract and research capabilities as native tools for a given agent framework.

Packages

Package	Published	Description
@tabstack/integrations-core	No (private)	Shared client, Zod schemas, tool bodies, error mapping, and types. Framework-agnostic.
@tabstack/langchain	Yes	LangChain.js adapter. Wraps core schemas and bodies in tool() from @langchain/core.
@tabstack/ai	Yes	Vercel AI SDK adapter. Wraps core schemas and bodies in tool() from ai.
@tabstack/eve	Yes	eve adapter. Wraps core schemas and bodies in defineTool() from eve/tools.
@tabstack/mastra	Yes	Mastra adapter. Wraps core schemas and bodies in createTool() from @mastra/core/tools.
@tabstack/claude-agent	Yes	Claude Agent SDK adapter. Wraps core schemas and bodies in tool() + createSdkMcpServer() from @anthropic-ai/claude-agent-sdk.
@tabstack/openai-agents	Yes	OpenAI Agents SDK adapter. Wraps core schemas and bodies in tool() from @openai/agents.
@tabstack/llamaindex	Yes	LlamaIndex.TS adapter. Wraps core schemas and bodies in tool() from llamaindex.
@tabstack/pi-agent	Yes	Pi coding agent extension. Registers core schemas and bodies via pi.registerTool() from @earendil-works/pi-coding-agent.

Design

• The SDK does the work. No HTTP, auth, or retry is reimplemented. Every network call goes through the SDK client.
• core knows no framework. It imports neither @langchain/* nor ai. Framework dependencies live only inside their own adapter package.
• core is the single source of truth. Tool names, descriptions, and Zod schemas are defined once and consumed by every adapter, keeping the cross-language and cross-framework story in lockstep with the shipped Python langchain-tabstack package.
• Adapters are thin shells. A new integration is a new folder under packages/ that imports core and writes an adapter. No speculative generic plugin abstraction.

Scope

@langchain/tabstack (the official LangChain-owned listing) is not built here. That is a separate contribution into the langchainjs repository and is out of scope for this monorepo.

Development

pnpm install
pnpm build      # build every package (tsup, dual ESM/CJS + .d.ts)
pnpm typecheck  # tsc --noEmit per package
pnpm test       # vitest per package
pnpm lint       # biome

Releases are driven by GitHub Release tags (langchain-vX.Y.Z, ai-vX.Y.Z, eve-vX.Y.Z, mastra-vX.Y.Z, claude-vX.Y.Z, openai-vX.Y.Z, llama-vX.Y.Z, pi-vX.Y.Z, or vX.Y.Z for all). core is private and excluded from release. See PUBLISHING.md.

Tooling choices

• tsup for builds: zero-config dual ESM/CJS output with .d.ts generation via a tiny per-package config. Mature and the de facto standard for small TypeScript libraries.
• biome for lint + format: one fast binary, one root config, no ESLint/Prettier plugin sprawl.
• vitest for tests; releases are cut from GitHub Release tags (see PUBLISHING.md).