Hey,i'mUdit

An MCP server for equity research and human-approved trading on NSE/BSE - a multi-agent trading system Claude and ChatGPT drive directly, with the host model reasoning while the server stays deterministic.

• 19 tools, 3 resources and 3 prompts - no model call and no model key on the server, which hands back indicator-level reasoning traces instead.
• Human-in-the-loop on every simulated trade - two-step orders (a preview, then a single-use expiring token) that pause for explicit approve/reject, and exactly one tool can place a trade.
• Caps, daily limits, kill switch and trading mode are enforced in the service layer, never from tool arguments - so no host call can bypass them.
• Hash-chained audit log and an AES-256-GCM encrypted credential vault.
• Replaced an earlier LangGraph supervisor CLI once the host model made it redundant - that CLI ran on ~10× fewer tokens per run (≈60K → ≈6K), traced end-to-end in LangSmith.

Document Q&A with no vector database and no embeddings - a four-stage query pipeline built on top of VectifyAI's open-source PageIndex TOC-tree indexer, with the upstream indexing kept intact.

• A four-stage query pipeline - Librarian, Navigator, Reader, Generator - carries a question from the TOC tree to a cited answer, with human confirm / override / skip at every stage.
• An LLM selects the pages and states its reasoning - dropping the embedding model and vector store entirely, and with them a whole layer of infra, cost and setup.
• Every answer comes back with page-level citations or an explicit refusal, and malformed model output is parsed defensively rather than trusted.
• All model calls routed through LiteLLM to NVIDIA NIM.
• Evaluation docs state what is measured - indexing accuracy through upstream's verify_toc gate - and what is not yet: recall@k, faithfulness, latency.

A multi-agent research engine that simulates a research team - parallel analyst personas run source-grounded interviews, synthesised into cited reports.

• Parallel analyst personas run concurrent, source-grounded interviews via a map-reduce flow.
• Stateful LangGraph runs with SQLite checkpointing - fully resumable from the last step.
• Every answer carries inline citations (Wikipedia / DuckDuckGo) - no hallucination.