Show HN: LangAlpha – What if Claude Code were built for Wall Street

We redesigned MCP and workspaces for agents working with financial data.

• MCP tools are too token-expensive for large-scale financial data. For example, a single call for 5 years of daily prices can put tens of thousands of tokens into the context window.
• When data vendors bundle dozens of tools into one MCP server, just the schemas can eat up 50k+ tokens.
• To avoid this, at workspace initialization we auto-generate typed Python modules from the MCP schemas and load them into the sandbox, and the agent imports them like ordinary libraries.
• The prompt only keeps a one-line summary per server. Whether a server has 3 tools or 30 tools, the prompt cost stays the same.
• This approach isn't finance-specific—it can be applied to any MCP server.

Making research continue across sessions is also a core piece. In investment research, work doesn't end with a single output—analysis keeps building, like updating a model when earnings come out or rerunning comps when a competitor reports earnings.

• Existing agents treat the job as done once a PDF or spreadsheet is produced, but in investment research that point is just the starting line.
• To solve the problem of files not naturally carrying over between sessions—forcing you to re-paste context every time—we built everything around workspaces, each mapped to a persistent sandbox.
• Each workspace has its own agent-dedicated memory file and file index, and this information is re-read before every LLM call.
• Even if you come back a week later in a new thread, the agent picks up where it left off and resumes work immediately.

We also inject financial domain context into the agent, much like codebase context.

• portfolio
• watchlist
• risk tolerance
• financial data sources

Some existing AI investment platforms offer this too, but not at the level a proper agent harness can deliver, so we built this system ourselves and released it as open source.