Le Chat: Custom MCP Connectors, Memories

Original Article ↗ Hacker News Discussion ↗

MCP Connectors and “Secure” Positioning

  • Announcement of 20+ “secure connectors” prompts questions about what “secure” means beyond admin control over connector access and on-behalf authentication.
  • Some wonder what concrete capabilities Stripe/PayPal MCPs provide (e.g. transaction search, balances, fees, FX rates).
  • A third-party dev asks how to get an open-source, multi-protocol file MCP (FTP/S3/SMB/etc.) listed in Mistral’s directory.

Model Speed, Cost, and Practical Quality

  • Several users report Mistral models as extremely fast and cheap, especially for summarization and high-volume pipelines.
  • Others say they were underwhelmed and see “Made in EU” as the main differentiator.
  • One detailed comparison: switching from gpt‑4.1‑mini/5‑mini to mistral‑medium yielded much better formatting adherence and ~10x speed, at similar cost, with occasional “harder” failures (random characters/backticks).
  • Some find Mistral weaker for factual QA/general knowledge and tool-heavy workflows, and not on par with frontier models (e.g. GPT‑5 Pro, high reasoning).

EU Origin, GDPR, and Data Governance

  • Strong thread around “European” as a selling point: GDPR compliance, lower geopolitical risk, and preference for non‑US providers handling PII.
  • Questions raised about how the new “memory” feature handles deletion and subject access requests; answers emphasize user responsibility and uncertainty about robust GDPR workflows for LLM memories.
  • Note that some fast services and image generation may still run in US data centers.

Developer Experiences and Structured Output

  • Multiple comments about dealing with LLMs surrounding JSON in markdown code fences or inserting stray characters; many handle this with regex post‑processing, schema-enforced inference, or tool/API-level schema constraints.
  • Discussion of techniques like prefilling outputs, retrying on schema failure, and grammar/regex-constrained decoding.

Market Position, Funding, and Competition

  • Mistral’s ~$14B valuation is viewed as low relative to US peers; some think this is an opportunity, others doubt long‑term survival against better-funded US companies.
  • Supporters highlight open-weight releases, on‑prem deployment help, and competitive pricing; skeptics say they still trail OpenAI/Anthropic/Google at the frontier.
  • Concern that successful European AI firms may eventually be acquired by US giants, though some believe France would block that on national-interest grounds.

European Tech, Politics, and Capital

  • Long tangent on why Europe lags in each tech wave: theories include fragmented markets, pension capital not flowing into VC, weaker incentives for extreme wealth creation, and stronger US/China internal markets.
  • Counterpoints argue that market size and policy, not “culture,” are primary; comparisons to China’s state-driven yet massive tech ecosystem.

Ecosystem, Integrations, and Product Gaps

  • Proton’s Lumo chat uses self-hosted Mistral Small along with other OSS models; seen as a privacy-friendly option.
  • Some ask why use Mistral MCPs instead of official vendor MCPs to avoid granting Mistral extra access.
  • Missing pieces noted: no desktop Le Chat client, remote-only connectors harder to use with local resources, and models struggling with multi-tool calls.

Design and Miscellaneous Notes

  • Several people like Le Chat’s visual design and branding; font choice (Arial) is debated.
  • Minor technical nit: Le Chat reportedly identifies itself as python-httpx/0.28.1 rather than a custom user agent.