Skip to main content

Tinfoil Provider Guide

Tinfoil is a Tier-2 catalog provider: its integration is one JSON file (src/lib/providers/catalog/tinfoil.json) rather than hand-written code. That file is the source of truth for this page.

Verification status: this entry is docs- and roster-verified, not yet live-verified. The fields below come from Tinfoil's own public pages and two unauthenticated GET requests (/v1/models and /v1/chat/completions) — no account was created and no API key was used to build it. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts​

  • Provider id: tinfoil (no alias)
  • Protocol: OpenAI-compatible (/chat/completions). The homepage says "Our API is OpenAI-compatible, open-source, and fully verifiable." (tinfoil.sh) and the chat models page says "Chat models use the OpenAI chat completions API." (docs)
  • Base URL: https://inference.tinfoil.sh/v1
  • Default model: kimi-k3
  • Models in catalog: 7 (the seven entries whose roster type is chat, taken from a 17-entry roster)
  • Streaming: supported — the structured outputs guide shows a request with stream=True that reads chunk.choices[0].delta.content, under the heading "Enable Streaming for Large Responses" (docs)
  • Tool calling: supported (true) — the tool-calling guide documents function calling (docs), and the roster's tool_calling field is true on the seven catalog models
  • Tools while streaming: not declared (false) — no combined tools-and-streaming request was sent (no credentials were used)
  • Structured output: supported — the structured outputs guide documents a response_format of type json_schema (docs), and the seven chat model cards carry a "Structured Outputs" line (docs)
  • Structured output + tools together: not declared (false) — no combined probe was run without credentials
  • Embeddings: not declared (false) on this catalog entry. The roster lists nomic-embed-text with type embedding (docs); it is not one of the seven catalog models
  • Thinking: declared (true) — the reasoning guide documents a reasoning_effort request parameter and says the model's thinking is returned in the reasoning field of the response message (docs); the accepted values differ by model (see Reasoning effort)
  • Vision: four catalog models (kimi-k3, gemma4-31b, glm-5-3-flash, deepseek-v4-1-flash) — see Models
  • Billing: free-with-card — see Billing below
  • Key format: none validated (apiKeyFormat is null)
  • Connection security: NeuroLink calls the endpoint directly over HTTPS — see Direct HTTPS access

Quick Start​

1. Get an API key​

  1. Visit: https://dash.tinfoil.sh?tab=api-keys and sign in or sign up — https://docs.tinfoil.sh/get-api-key says "Create an account with your email or one of the social connections."
  2. On the dashboard Products page, click Activate on the Private Inference card. Per https://docs.tinfoil.sh/get-api-key, if no payment method is on file you are redirected to a checkout page to enter your card details, and "You’re only charged based on usage."
  3. The activation dialog creates a "Default API key" and, per the same page, "adds $1 in API credits to your account"; additional keys are created from API Keys in the sidebar
  4. Set TINFOIL_API_KEY in your .env file (the examples on https://docs.tinfoil.sh/quickstart read TINFOIL_API_KEY)
  5. NeuroLink sends requests straight to https://inference.tinfoil.sh/v1. https://docs.tinfoil.sh/sdk/direct-api-access says "but this is not recommended in production" and "Direct access skips the automatic attestation verification our SDKs perform". https://docs.tinfoil.sh/local-proxy/cli documents a local proxy that verifies enclave attestations and serves http://127.0.0.1:3301/v1 for OpenAI-compatible clients (not exercised with NeuroLink)

2. Configure​

export TINFOIL_API_KEY=your-api-key
export TINFOIL_MODEL=kimi-k3   # optional — overrides the default model
export TINFOIL_BASE_URL=https://inference.tinfoil.sh/v1   # optional — proxy or self-hosted gateway

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "tinfoil",
  model: "kimi-k3",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider tinfoil

Per-request credentials work as they do for the other catalog providers:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "tinfoil",
  credentials: { tinfoil: { apiKey: process.env.TINFOIL_API_KEY } },
});

Direct HTTPS access​

The catalog's base URL, https://inference.tinfoil.sh/v1, is the endpoint the Tinfoil docs call "direct" access. The docs page for it (Direct API Access) says "but this is not recommended in production" and "Direct access skips the automatic attestation verification our SDKs perform, so you have no cryptographic proof your data is processed in a secure enclave and no protection against man-in-the-middle attacks." It also says "We maintain direct HTTPS compatibility for debugging and simple testing (e.g., inference requests via cURL)."

The coding-agents tutorial says "OpenAI-compatible clients don’t perform connection-time attestation themselves." (docs) and the Tinfoil Proxy CLI page describes a local proxy that listens on http://127.0.0.1:3301, verifies the enclave's attestation, and takes that address as the base URL for any OpenAI-compatible client. Pointing TINFOIL_BASE_URL at it follows those docs; NeuroLink through that proxy was not exercised.

Billing​

Tinfoil's key guide (docs) describes the flow: sign up, click Activate on the Private Inference card, enter card details at checkout when no payment method is on file ("You’re only charged based on usage."), and receive a "Default API key" from a dialog that "adds $1 in API credits to your account". The homepage labels the API "Usage-based pricing" (tinfoil.sh). The pricing page's Private Chat card lists $20 per month (pricing); Private Chat and Private Inference are separate entries in the site's Products menu.

The entry records free-with-card, the value among the schema's three (free-tier, free-with-card, no-free-tier) that matches a card at checkout plus a stated credit. Per-token prices are in the Models table.

Models​

ModelContextVisionPrice per 1M tokens in / out (cached in)Notes
kimi-k3 (default)262,144yes4 / 20 (0.8)NeuroLink default. Roster description: "Moonshot's flagship multimodal MoE model with hybrid attention for long-context reasoning, coding, and agentic workflows".
gpt-oss-120b131,072no0.15 / 0.6Fallback. The first inference example on the quickstart uses this model ID.
gemma4-31b262,144yes0.4 / 1Fallback. Strengths value on the chat models page: "Built-in thinking mode, image understanding, native function calling, multilingual support for 35+ languages".
llama3-3-70b131,072no1.75 / 2.75Fallback. Roster description: "High-performance multilingual language model optimized for dialogue".
glm-5-31,048,576no1.8 / 5.75 (0.45)Roster description: "Z.ai's latest flagship for agentic coding and long-horizon tasks, with a 1M-token context window".
glm-5-3-flash1,048,576yes0.4 / 1.25 (0.1)Roster description: "Z.AI's fast multimodal MoE model with a 1M-token context window, image input, reasoning, and tool calling". Roster field experimental is true; recorded as preview.
deepseek-v4-1-flash1,048,576yes0.65 / 1.45 (0.13)Roster field experimental is true; the chat models page shows the warning "this model is not yet fully productionized. Capacity and performance may change while we finish rolling it out." Recorded as preview.

Context and prices are the context_window and pricing fields of an unauthenticated GET https://inference.tinfoil.sh/v1/models call made 2026-09-29 (17 entries; the seven with type chat are the seven models above). The chat models page gives the Context values in rounded form: "256K tokens" for kimi-k3 and gemma4-31b, "131K tokens" for gpt-oss-120b, "128K tokens" for llama3-3-70b, and "1M tokens" for glm-5-3, glm-5-3-flash and deepseek-v4-1-flash. maxOutputTokens is unset in the catalog. models.defaultContextWindow (131,072) and models.defaultMaxOutputTokens (4,096) are placeholders the vendor does not publish.

Status: production is recorded for kimi-k3, gpt-oss-120b, gemma4-31b, llama3-3-70b and glm-5-3, whose roster entries do not set experimental. preview is recorded for glm-5-3-flash and deepseek-v4-1-flash, whose roster entries set experimental to true.

Left out of the catalog: the ten roster entries whose type is not chat — gpt-oss-safeguard-120b (safety), pii-filter (safety), nomic-embed-text (embedding), doc-upload (document), voxtral-small-24b, whisper-large-v3-turbo, voxtral-mini-4b-realtime (audio), qwen3-tts, voxtral-tts (tts) and websearch (tool). gpt-oss-safeguard-120b, voxtral-small-24b and websearch list /v1/chat/completions among their endpoints.

Fallback order when the default is unavailable: gpt-oss-120b → gemma4-31b → llama3-3-70b. The runtime fallback model name the loader derives (fallbacks[1]) is gemma4-31b.

Vision: vision: true for kimi-k3, gemma4-31b, glm-5-3-flash and deepseek-v4-1-flash. The image processing guide names those four as supporting image inputs, the vision models page lists them, and the roster reports multimodal: true for them. For gpt-oss-120b and llama3-3-70b the image processing guide says "Other models (Llama, GPT-OSS) are text-only and cannot process images."; glm-5-3 has multimodal: false on the roster. No image request was sent.

Reasoning effort​

Accepted reasoning_effort values per model, from the reasoning guide:

ModelSupported values
kimi-k3low, high, max
deepseek-v4-1-flashnone, low, high, xhigh, max
glm-5-3low, high, max
glm-5-3-flashlow, high, max
gemma4-31bnone, minimal, low, medium, high, xhigh, max
gpt-oss-120blow, medium, high

The same page says Kimi K3, GLM-5.3 and GLM-5.3 Flash always use reasoning and default to max, and that sending an unsupported value returns a 400 error. The chat models page states that for GLM-5.3 and GLM-5.3 Flash reasoning tokens count toward max_tokens, so structured outputs and strict tool calls at the default effort need a generous max_tokens (around 20K) or a lower effort.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for Tinfoil — docs- and roster-verified, not live-verified:

ProbeResult
Rosterunauthenticated GET https://inference.tinfoil.sh/v1/models, HTTP 200, 17 entries (7 with type chat), 2026-09-29 — a second GET the same day returned an identical body; no key used
Billingpublic page https://docs.tinfoil.sh/get-api-key, 2026-09-29: card details at checkout when no payment method is on file, "You’re only charged based on usage.", and the activation dialog "adds $1 in API credits to your account"
Auth-failure shapeunauthenticated GET https://inference.tinfoil.sh/v1/chat/completions, 2026-09-29, returned HTTP 400 with {"error":{"message":"Invalid request body: EOF.","type":"invalid_request_error","param":null,"code":"invalid_json"}}; errorRules therefore use the codes and messages documented at https://docs.tinfoil.sh/sdk/error-handling (invalid_api_key, insufficient_quota on 402 and 429)
Tools / structured output / streamdocumented at https://docs.tinfoil.sh/guides/tool-calling and https://docs.tinfoil.sh/guides/structured-outputs; none was exercised live, no POST was sent, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=tinfoil with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Troubleshooting​

SymptomCauseFix
HTTP 402 (insufficient_quota)The docs list message "You exceeded your current quota, please check your plan and billing details." and cause "The account or organization has no remaining balance or an inactive subscription"Billing in the Tinfoil dashboard sidebar (docs)
HTTP 429 (insufficient_quota)A per-key cap. The docs list the messages "Spend limit reached for this API key.", "Input token limit reached for this API key." and "Output token limit reached for this API key."The docs describe a 429 from a per-key cap as not transient, so retrying will not help, and say to raise or clear the cap with Update API Key or use a different key (docs)
HTTP 400 when a reasoning_effort is sentThe reasoning guide says accepted values differ by model and an unsupported value returns a 400 errorUse a value from the per-model table in Reasoning effort
Structured output ignored with tools attachedstructuredOutputWithTools is false on this entry — untested combinationNeuroLink omits response_format automatically whenever tools are present, before sending

See also​