TypeScript-first AI SDK | 13+ providers, one API | MCP-native | Open source by @Juspay
Global
Joined January 2026
- Tweets167
- Following28
- Followers3
- Likes11
Pinned Tweet
A year ago, nobody talked about MCP.
Today: 97M monthly downloads. Every major AI lab supports it. The Linux Foundation governs it.
Here's why MCP matters for every developer building AI agents:
🧵
We flipped one answer in a copy of our results: 4,667 → 4,666. Our checker caught it. ScaleDown Classify via Juspay's AI harness (NeuroLink), one run each, 8,580 questions: BANKING77 77.1%, CLINC150+OOS 84.9%. $2.25 at list price. Count it, break it, bill it. Not a ranking.
OpenAI's GPT-6 Luna Decisions, accuracy: BANKING77 77%, CLINC150+OOS 88%. OpenRouter's served route (not OpenAI's own API), via NeuroLink's typesafe provider pointed at OpenRouter. One run each, a failed answer counted wrong. XOR is Juspay's own model.
OpenAI's guide: Decisions bills input tokens only, $0.10 per 1M. Every choice request carries its options. NeuroLink, Juspay's AI harness: BANKING77 requests (77 options) averaged 995 input tokens at Perplexity, 1,866 at Cloudflare.
We counted on Oct 8, 2026: 14 decision models from 10 makers in OpenRouter's list or rankings, 13 with a live endpoint. In its 30-day category rankings (usage data through Oct 7, 2026), Jev is first in all 23 categories by share of spend and of tokens.
NeuroLink, Juspay's AI harness: provider typesafe, TYPESAFE_BASE_URL at openrouter.ai/api, model perplexity/pplx-decider-v1.1-27b. 200 BANKING77 rows matched our Perplexity API run (expected: OpenRouter lists only Perplexity). Check cost: $0.0041.
Perplexity Decider v1.1 is now on OpenRouter
@perplexity_ai's open-weights multimodal decision model takes text, JSON, or images and returns typed answers with probabilities
$0.02/M input. Output is free
openrouter.ai/perplexity/ppl…
NeuroLink 12.49.4's proxy keeps every system block Claude Code sends exactly as sent, cache markers included. In our September log, from an earlier build, 95.6% of Sonnet 5 prompt tokens were cache reads, the kind Sonnet 5.5 now prices at half.
Two v1.1 runs via NeuroLink, Juspay's AI harness. Cost at $0.02/M input tokens: BANKING77 $0.061, CLINC150+OOS $0.182, two of Decision Index 0.3's 37 public benchmarks. Same requests to v1, one run each. Correct: v1.1 81.0% and 91.6%, v1 81.2% and 91.7%.
pplx-decider-v1.1-27b, our updated open weights multimodal decision model, is now available.
It scores the highest on the new @huggingface Decision Index 0.3 benchmark. pplx-decider-v1.1-27b costs half as much as v1, at $0.02 per million input tokens.
huggingface.co/spaces/multim…
Clef-flash vs Clef via NeuroLink, one run each. BANKING77 accuracy: 90.9% vs 94.1%. CLINC150+OOS: 63.1% vs 97.3%. It caught 991 of 1,000 out-of-scope requests, but 1,796 of its 2,019 wrong in-scope answers were "out of scope". Two of Decision Index 0.3's 37 public benchmarks.
Four decision models, two public tests, via NeuroLink (Juspay's AI harness). Plain accuracy, not the JevBench Score; averaged over runs. JevBench (193 of 231 rows): XOR 92%, Clef 90%, Jev 89%, Laya 60%. Typed Decisions: Jev 74%, Clef 72%, XOR 67%; Laya 77% (fine-tuned on it).
What we're doing: Laya via NeuroLink (Juspay's AI harness), one duplicate-charge ticket in eight languages, each routed to a team by decide(). What it achieved: all eight went to billing, each call under 400 ms; English to the English checkpoint, the rest to the multilingual one.
Longer version of this, on LinkedIn: what changed in NeuroLink's decide() when a second, self-hosted provider (Laya) joined the first hosted one (Jev).
linkedin.com/feed/update/urn…
decide() has a second provider: Laya, Convai Innovations' open-weights System One model, next to TypeSafe's Jev. Same call site, either provider — the real test of whether the abstraction generalized.blog.neurolink.ink/posts/two…
generate. stream. decide.
NeuroLink now treats a decision as a third inference type — ~400ms, ~$0.00002, no text output.
Model routing, context pruning, MCP server selection, RAG planning.
blog.neurolink.ink/posts/gen…
Longer write-up on LinkedIn — the full cost breakdown, and the two bugs that shipped silently because they produced plausible output:
linkedin.com/feed/update/urn…
5 TypeScript patterns that will save your production AI system:
(Learned from processing millions of AI requests at Juspay)
🧵
3. Generic types for provider-agnostic functions.
function generate<T extends Provider>(
config: ProviderConfig<T>
): Promise<TypedResponse<T>>
Your code stays type-safe even when switching between 13+ providers.
This is what NeuroLink does under the hood.
4. AsyncIterator for streaming.
5. Branded types for API keys (never mix up provider keys).
Full implementation patterns in the NeuroLink source:
github.com/juspay/neurolink
TypeScript + AI = fewer bugs, better DX, safer production.
What patterns would you add?
