Skip to main content
Hindsight MCP 2.0 is the first stable release of our Model Context Protocol server. It moves authentication to OAuth, replaces the old single-agent entry point with direct tool calls, and ships a round of reliability fixes.
This is a breaking change. Connections built against the pre-2.0 server need to be reconnected. The upside: 2.0 is now our stable contract — tool names, auth, and behavior won’t shift underneath you the way the preview server could. See What you need to do below.

​
OAuth is now the primary sign-in

The MCP server now authenticates with OAuth as the primary flow. When you install the Hindsight connector, Claude opens your browser, you sign in to Hindsight, and you pick the workspace to use — no copying API keys into a config file. What changed under the hood: access is granted through Clerk-issued OAuth tokens, verified per request, with automatic token refresh. The server advertises an OAuth-protected-resource endpoint so compatible clients can discover and complete the sign-in flow on their own. What it means for you:
  • Per-user access, not a shared key. Each person signs in as themselves, and access is scoped to the workspace they select — better security and cleaner usage attribution.
  • Simpler setup. The one-click connector handles auth end-to-end; there’s nothing to paste or rotate.
  • API keys still work for manual/headless setups. Bearer-key configs (Authorization: Bearer YOUR_API_KEY) remain supported as a fallback — useful for servers and CI. See the MCP setup guide.

​
Direct tool calls

The biggest architectural change: the server now exposes Hindsight’s capabilities as individual tools your assistant calls directly, instead of routing everything through a single orchestrating agent that spun up sub-agents behind the scenes. Your assistant now calls focused tools like fetch_hindsight_context, select_deal, get_deal, search_deal_documents, create_dataset, and analyze_dealgrid_subset itself — and chains them as it sees fit. What it means for you:
  • Faster responses. Skipping the extra agent-orchestration layer cuts latency — simple questions don’t pay for a full sub-agent spin-up.
  • Fine-grained control. You can see exactly which tool runs at each step, approve or deny individual calls, and disable tools you don’t want exposed. Nothing happens inside an opaque agent loop.
  • Better composability. Your own model decides how to combine Hindsight’s tools with everything else it can reach, which is what makes the server a clean building block for custom agents and skills.
The canonical tool list is generated from the live server. For the current surface and how the tools fit together, see the Platform Overview and MCP docs.

​
Fixes & reliability

  • Cross-origin support — added CORS handling (including preflight OPTIONS) so browser-based clients like Claude and ChatGPT on the web connect cleanly.
  • OAuth resource resolution — the protected-resource URL is now derived from the deployment’s site URL, fixing discovery in production.
  • Tool-result parsing — more robust extraction of result summaries so responses render reliably across clients.
  • Tool-call rendering — corrected how in-progress tool calls display during a response.

​
What you need to do

If you connected to the preview server, reconnect against 2.0:
  1. Reinstall the connector — download the latest hindsight.mcpb and install it in Claude. You’ll be prompted to sign in with OAuth and select your workspace.
  2. Or update your manual config to point at https://app.usehindsight.com/api/mcp. API-key Bearer auth still works for manual setups.
  3. Re-check tool permissions in your client — with direct tool calls, you control which tools are enabled.
Full instructions live in the MCP setup guide.

​
Coming soon

Web MCP as an official Claude connector. We’re working with Anthropic to offer Hindsight as a one-click connector inside Claude on the web — no local bundle to install. The OAuth flow in this release is the groundwork for it.