# Stacklink

> Stacklink is the integration, knowledge, and execution boundary for AI agents. It owns external provider accounts, tool discovery and execution, verified knowledge packs with evidence receipts, triggers and webhooks, and a remote workbench — so consumer apps keep only their own product reasoning and UX.

## Instructions

- Choose one onboarding path: connect an AI app over the stable project MCP resource and Stacklink OAuth, or add the public `@stacklink/sdk` / `stacklink-sdk` package to code you own.
- Never place a Stacklink API key or runtime session ID in an MCP app installation URL. The client completes OAuth; authenticated `initialize` creates the internal session and `tools/list` proves the connection.
- For SDK applications, create a project API key and an explicit end-user session with `stacklink.create({ tenantId, userId })` in TypeScript or `sessions.create_bound(...)` in Python. Close the session in `finally`.
- Pass Stacklink executable tools or MCP configuration to Vercel AI SDK, OpenAI Agents SDK, or Claude Agent SDK. The selected harness owns the model call, tool loop, and final answer.
- The execution loop is: `STACKLINK_SEARCH_TOOLS` (discover by intent) -> `STACKLINK_GET_TOOL_SCHEMAS` (fetch exact JSON schema) -> `STACKLINK_EXECUTE_TOOL` / `STACKLINK_MULTI_EXECUTE_TOOL` (run). Resolve credentials by `accountId`/`accountAlias`; raw credentials never reach the model.
- To connect an end-user account, call connection authorize, send the user to the returned `redirectUrl`, then poll the connection request until it completes.
- For knowledge packs: prefer structured tools (e.g. `GST_CALCULATE_DUE_DATE`) for deterministic math; for prose use `KNOWLEDGE_SEARCH`, preserve every returned `evidenceId`, and call `KNOWLEDGE_VERIFY_ANSWER` before you present a final answer. Honor `agentGuidance` and `currentnessState`.
- Timestamps are India time (`Asia/Kolkata`, `+05:30`).

## Get started

- [Connect your agent](/llms.mdx/docs/connect-your-agent/content.md): Add Stacklink to Cursor, VS Code, Claude, Codex, or another remote MCP client.
- [Overview](/llms.mdx/docs/content.md): Stacklink is the integration, knowledge, and execution boundary for AI agents.
- [Quickstart](/llms.mdx/docs/quickstart/content.md): Choose MCP for an AI app or the SDK for code you own.

## Concepts

- [Architecture](/llms.mdx/docs/architecture/content.md): How a request flows, the subsystems, and the deployment shape.
- [The boundary](/llms.mdx/docs/boundary/content.md): What Stacklink owns, and what your app owns.
- [Glossary](/llms.mdx/docs/glossary/content.md): The Stacklink vocabulary, defined once.
- [MCP server](/llms.mdx/docs/mcp/content.md): Stable project resources, OAuth, internal sessions, tools, revocation, and advanced API-key access.
- [Projects, tenants & users](/llms.mdx/docs/projects/content.md): The identity model, plus your API keys and scopes.
- [Sessions](/llms.mdx/docs/sessions/content.md): The unit of scope — tools, accounts, knowledge, and the MCP surface.
- [Tools & discovery](/llms.mdx/docs/tools-and-discovery/content.md): Toolkits, tools, and the search → schema → execute loop.

## Connections & auth

- [Auth](/llms.mdx/docs/auth/content.md): Auth configs, auth modes, and managed vs custom.
- [Connect an end-user account](/llms.mdx/docs/connect-an-account/content.md): Run a tenant through the hosted connect flow and reuse the account.
- [Connections](/llms.mdx/docs/connections/content.md): Connected accounts — link a provider account for your end-user.

## Triggers & webhooks

- [Triggers](/llms.mdx/docs/triggers/content.md): Subscribe a tenant to provider events, by webhook or polling.
- [Webhooks](/llms.mdx/docs/webhooks/content.md): Receive signed, normalized event deliveries and verify them.

## Executions & workbench

- [Executions & control](/llms.mdx/docs/executions/content.md): Every run is durable — list it, replay its events, wait on it, cancel it.
- [Workbench](/llms.mdx/docs/workbench/content.md): A safe scratch computer where your agent can run code, crunch data, and make files.

## Verified Knowledge

- [Evidence receipts](/llms.mdx/docs/knowledge/evidence-receipts/content.md): Cite, verify, and prove a compliance answer.
- [How packs work](/llms.mdx/docs/knowledge/how-it-works/content.md): The data model and retrieval behind verified knowledge.
- [Verified Knowledge](/llms.mdx/docs/knowledge/content.md): Citation-backed compliance knowledge an agent can prove, not just recall.
- [Structured tools](/llms.mdx/docs/knowledge/structured-tools/content.md): Deterministic compliance calculations, each backed by evidence links.

## Platform

- [Billing & usage](/llms.mdx/docs/billing/content.md): Simple, metered pricing — most things are standard ops, heavy compute is premium.
- [Errors & idempotency](/llms.mdx/docs/errors/content.md): The error envelope, common codes, and safe retries.
- [Observability](/llms.mdx/docs/observability/content.md): Two log levels, queryable over the API, retained per plan.

## Reference

- [Configuration](/llms.mdx/docs/reference/configuration/content.md): Environment variables, grouped by subsystem.
- [Reference](/llms.mdx/docs/reference/content.md): Meta-tools, providers, SDKs, webhook events, scopes, and configuration.
- [Meta-tools](/llms.mdx/docs/reference/meta-tools/content.md): The model-callable tool surface exposed by every session.
- [Providers](/llms.mdx/docs/reference/providers/content.md): The integration catalog — auth modes, capabilities, and trigger support.
- [API-key scopes](/llms.mdx/docs/reference/scopes/content.md): Exactly what each key can do.
- [Python SDK](/llms.mdx/docs/reference/sdk-python/content.md): Certified OpenAI Agents SDK and Claude Agent SDK recipes.
- [TypeScript SDK](/llms.mdx/docs/reference/sdk-typescript/content.md): Certified Vercel AI SDK, OpenAI Agents SDK, and Claude Agent SDK recipes.
- [Webhook events](/llms.mdx/docs/reference/webhook-events/content.md): The delivery envelope, signature headers, and event types.