Developer Arhancanli has released canli-mcp, a Model Context Protocol (MCP) server that consolidates 306 financial tools into a single, free, MIT-licensed package. Designed for Claude, Cursor, and other MCP clients, the tool runs locally on the user's machine and requires no API keys for the vast majority of its functions. The architecture exposes just three meta-toolsβ€”find_tool, describe_tool, and run_toolβ€”to keep context windows manageable while providing access to deep financial data.

Architectural Efficiency and Token Costs

The core engineering challenge with large tool sets is context bloat. canli-mcp addresses this by sending only 859 tokens of tool definitions per request, compared to OpenBB's 432,001 tokens (or 2,171 in discovery mode) and EdgarTools' 3,801 tokens. Even when installed as seven separate servers, the Canli ecosystem sends 34,508 tokens, whereas the unified canli-mcp reduces this to 1,109 tokens. This drastic reduction in overhead makes high-fidelity financial analysis viable for models with limited context windows.

Benchmark Performance Against Rivals

In head-to-head testing using gpt-5.4-mini, canli-mcp answered 22 of 28 questions correctly (79%), outperforming the best rival setup of EdgarTools plus Yahoo Finance, which achieved 54% (15 of 28). The test covered 53 finance question types, with canli-mcp providing keyless tools for 49 of them, versus 39 for OpenBB and only 11 for EdgarTools. The author notes that the test set was written by the developer and covers areas the server was specifically built for, with a held-out set still pending.

Data Integrity and Search Logic Fixes

The release highlights significant improvements in search accuracy and data validation. A critical bug where the stemmer converted "rates" to "rat" prevented matches for "mortgage rates" and similar queries; this has been resolved. Additionally, ambiguous entity resolution was fixed to ensure "Facebook" maps to Meta and "JP Morgan" to the correct filer. The server now refuses requests it cannot answer with specific words, preventing errors like returning US CPI when asked for "silver price."

Privacy and Provenance

canli-mcp emphasizes local execution and data provenance. It connects only to public sources such as the SEC, Treasury, FRED, FINRA, Cboe, and Yahoo, sending no data to the developer and logging nothing locally. Market data responses include hashes of the exact bytes received, allowing users to verify figures later. An offline mode (CANLI_OFFLINE=1) restricts the server to packs that require no network access, further enhancing privacy.

Key Takeaways

  • canli-mcp reduces token overhead for 306 finance tools to 859 tokens per request, significantly lower than OpenBB and EdgarTools.
  • The server achieved a 79% accuracy rate in head-to-head testing against gpt-5.4-mini, surpassing competitors.
  • Search logic was refined to fix stemming errors and improve entity resolution for major companies and financial instruments.
  • The tool is free, MIT-licensed, and runs locally with no API keys required for most functions.

The Bottom Line

This is a pragmatic solution to the 'tool bloat' problem in agentic AI; by compressing 306 tools into a searchable interface with negligible token cost, canli-mcp makes deep financial analysis accessible without the massive context tax imposed by competitors.