Documentation

Quickstart

Get started with TemuClaude in under 5 minutes.

Option 1 — Use the playground (no installation):

Open the playground → — sign in, ask anything, and get a superior answer. 20 free queries/day.

Option 2 — OpenAI-compatible API access:

curl -X POST https://temuclaude.com/v1/chat/completions \
  -H "Authorization: Bearer tmc_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"model":"temuclaude","messages":[{"role":"user","content":"What is 9.9 vs 9.11?"}]}'

 # Response: OpenAI-compatible chat completion JSON

Data Privacy & Sovereignty

TemuClaude is built on a foundation of absolute data safety and developer trust:

• Zero Log Retention — We process all requests in-memory. Your raw query content and code outputs are never persisted to any databases or disk logs.
• API Key Encrypted Storage — All custom API keys and session credentials are encrypted in-transit and at-rest using AES-256 standard encryption.
• No Data Reselling — Your training vectors and query histories belong 100% to you and are never used to train internal models or sold to third-party resellers.

Architecture

TemuClaude is one endpoint that assigns each model an evidence-based capability role. When you ask a question, it:

1. Classifies your query (math, coding, reasoning, knowledge, creative, legal, health, agentic)
2. Estimates difficulty (trivial, medium, hard)
3. Routes to the best strategy:
   • Pro trivial → GLM-5.2 quality floor
   • Pro nontrivial → all available role specialists in parallel, followed by synthesis and verification
   • Lite nontrivial → parallel DeepSeek/Qwen drafts, Qwen synthesis, and Nemotron verification
4. Adapts step models for search, verification, consistency, QA gates, debate, and post-processing using telemetry when enough evidence exists
5. Returns one clean answer — orchestration is invisible but visible in the playground

TemuClaude Lite

Lite keeps the same authenticated Playground, task classification, answer trace, and safety policy as Pro, with a bounded quality ensemble designed for lower cost.

• DeepSeek V4 Flash is the default worker.
• Qwen 3.7 Plus handles hard structured reasoning and agentic work.
• Qwen 3.7 Plus handles vision, UI, long-context, and agentic work.
• DeepSeek and Qwen produce complementary drafts in parallel for every nontrivial request.
• Qwen 3.7 Plus synthesizes the strongest complete result.
• Nemotron 3 Ultra independently verifies every nontrivial result and risky or sampled trivial work.

Model Pool

TemuClaude has ten active routing roles. Every available Pro specialist participates in nontrivial work with a distinct role prompt.

Adaptive Pipeline (up to 10 layers for hard Pro)

Trivial and medium Pro queries take a fast, efficient path — a single specialist or light shepherding. For genuinely hard Pro queries, TemuClaude runs a bounded quality pipeline with up to 10 layers designed to improve completeness and catch errors. Stages marked research runtime belong to the Python engine / playground and do not run on every request:

1. Web Search — research runtime. Dynamic knowledge retrieval for real-time information query enhancement.
2. MoA 3-Layer Fusion — Layered peer reviews where multiple specialist models refine each other's outputs before aggregation.
3. Self-Consistency — Weighted voting frameworks selecting the most logically consistent path.
4. Code Review — Independent coding-agent and verifier models critique outputs to identify implementation defects and missing requirements.
5. Reflexion — Verbal correction logic to automatically re-evaluate reasoning steps when validation indicators flag errors.
6. Self-QA Gate — An independent verifier model (Nemotron 3 Ultra) scores answers on logical coherence, completeness, and goal alignment.
7. Consistency Verification — Independent checks look for contradictions in structured arguments.
8. Budget Forcing — research runtime. Adaptive token management extending system computation time on complex derivations.
9. Step-Level Verification — research runtime. Independent validation of intermediate logical statements.
10. Frontier Adjudication — Frontier models contribute proposals and return for corrective synthesis when QA is low.

3-Tier Routing

TemuClaude classifies queries dynamically using semantic analysis. Queries are mapped to the most efficient tier that can guarantee optimal response quality:

• Standard Tiers — High-speed, focused specialist engines for trivial and factual inquiries.
• Advanced Tiers — Domain-specific reasoning engines best suited for structural or complex contexts.
• Elite Tiers — Full multi-layer consensus loops with logical validation and parallel generation.

Step-Aware Model Router

Instead of static routing, TemuClaude continuously evaluates routing paths. The model selection varies dynamically across different steps within a single query response path, ensuring each logical task is handled by the most optimized resource.

Active Budget Controller

Our controller analyzes safety margins and reasoning path indicators to steer queries dynamically. It cuts off dead-end reasoning loops early and allocates computation resources precisely where they provide the greatest quality return.

MoA 3-Layer Fusion

For high-complexity tasks, TemuClaude runs a multi-stage consensus process. Specialist models independently generate draft solutions, exchange peer reviews to highlight blind spots and contradictions, and synthesize a single, optimized consensus response.

Self-Consistency

For quantitative and reasoning tasks, the system generates multiple candidate pathways. A weighted voting classifier selects the final path based on logic consistency, mitigating hallucination spikes.

Code Review

For hard Pro coding and math queries, an independent judge model reviews the generated code and reasoning for defects, missing requirements, and incorrect logic. Sandbox execution and Z3 constraint solving belong to the Python research runtime for hard queries, not the live gateway. When review flags an error, the reflexion loop rewrites and corrects the draft.

Self-QA Gate

For nontrivial queries, generated answers pass through an automated quality gate — scored by an independent verifier model (Nemotron 3 Ultra) — evaluating logical consistency, factual correctness, and goal alignment. If the thresholds are not met, the response is rejected and sent back to the pipeline for refinement.

Reflexion

When the quality check flags a discrepancy, the system generates detailed critique logs. The generator models use these critique logs to reformulate their response, correcting logical missteps.

Budget Forcing

To handle deep reasoning queries, TemuClaude can dynamically extend the thought process. This forces models to generate comprehensive step-by-step proofs instead of jumping directly to short, unverified conclusions.

Z3 Logical Verification

For hard Pro structural-logic questions, the Python engine translates key relationships into symbolic constraints and runs them through a Z3 logic solver to verify satisfiability. If any logical contradictions are detected, a multi-agent debate is triggered to resolve them. This runs in the research runtime for hard queries and falls back gracefully when Z3 is unavailable.

Frontier Participation

Available frontier models participate in the first panel for every nontrivial Pro request. When verification remains below the quality threshold, they receive the accumulated draft and feedback for a corrective adjudication pass.

Methodology

TemuClaude's benchmark scores are projected from:

1. Individual model scores from ArtificialAnalysis Intelligence Index
2. Proven improvement numbers from published research papers
3. Stack effects: each technique's proven gain applied to the baseline
4. Conservative estimates (lower bounds of published ranges)

Evaluation & Trust

We believe in rigorous scientific evaluations and absolute transparency. While the core orchestration weights and routing heuristics are proprietary to prevent direct duplication of our infrastructure, we provide extensive validation data:

• Independent Audits — Enterprise clients can request access to sandbox evaluation environments to run custom test suites.
• Benchmark Logs — Full transaction traces, model outputs, and logical consensus trees are logged during evaluations to verify reproducibility.
• Custom Verification — We support zero-knowledge auditing tools to verify that our routing runs exactly as documented without exposing proprietary models.

Projected Scores

TemuClaude against a leading frontier baseline across 8 benchmarks:

* Projected from research analysis. Live results pending ArtificialAnalysis verification.

REST API

Use the OpenAI-compatible endpoint for assistants and developer tools:

POST https://temuclaude.com/v1/chat/completions

Request:
{
  "model": "temuclaude",
  "messages": [
    {"role": "user", "content": "What is 9.9 vs 9.11?"}
  ]
}

Response: OpenAI-compatible chat completion JSON

Use model temuclaude. The server keeps model selection and orchestration internal.

Streaming

TemuClaude uses Server-Sent Events (SSE) for streaming responses:

• POST request with messages array
• SSE response — chunks stream as data: {chunk}
• Orchestration data — sent after all chunks, before [DONE]
• [DONE] — signals stream end

Orchestration Data

The final SSE event contains orchestration metadata showing exactly how the answer was built:

{
  "orchestration": {
    "taskType": "math",
    "tier": "hard",
    "models": [
      {"name": "glm-5.2", "response": "...", "latency": 3.2, "correct": true},
      {"name": "deepseek-v4-pro", "response": "...", "latency": 8.3, "correct": true},
      {"name": "gemini-3-flash", "response": "...", "latency": 13.9, "correct": true}
    ],
    "aggregator": "glm-5.2",
    "consensus": 3,
    "qaScore": 8,
    "codeVerified": true,
    "totalLatency": "48.2",
    "cost": "$0.015",
    "techniques": ["moa-3-layer", "cross-review", "structured-aggregation",
                   "self-consistency", "prm-weighted-voting", "code-verification",
                   "reflexion", "usva-4-rubric-qa", "s1-budget-forcing"]
  }
}

Authentication

API requests require a TemuClaude API key generated in the developer dashboard. Keys begin with tmc_; keep them server-side and revoke unused keys immediately.

curl -X POST https://temuclaude.com/v1/chat/completions \
  -H "Authorization: Bearer tmc_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"model":"temuclaude","messages":[{"role":"user","content":"Hello"}]}'

Use with Hermes Agent

Recommended for agent workflows: Hermes can use TemuClaude as its reasoning backend while Hermes provides its own tools, workspace access, skills, and approvals.

hermes model

# Choose: Custom endpoint (self-hosted / VLLM / etc.)
# Base URL: https://temuclaude.com/v1
# API key:  tmc_your_api_key
# Model:    temuclaude

# ~/.hermes/config.yaml
model:
  provider: custom
  base_url: https://temuclaude.com/v1
  api_key: tmc_your_api_key
  default: temuclaude

Rate Limits

Error Codes

Media Orchestration

TemuClaude also orchestrates media generation — images, video, text-to-speech, and music. A similar multi-stage pipeline (cache, intent, tier, parallel generation, judge, quality gate, reflexion, memory, return) — up to 10 stages for complex media requests. Each media type has its own model pool, routing logic, and quality gates.

Image Generation

3-tier routing with best-of-N generation and LLM judge:

Unique routing for special cases:

• Vector/SVG output → Recraft V3
• Text in images → FLUX-2 Flex, GPT Image 2
• Extreme aspect ratios → Nano Banana 2
• Character consistency → Nano Banana 2, GPT Image 2
• Multilingual text → Seedream 4.5

# Generate an image
curl -X POST https://temuclaude.com/api/media/generate \
  -H "Content-Type: application/json" \
  -d '{"type": "image", "prompt": "a cat on a windowsill", "tier": "standard"}'

Video Generation

3-tier routing with async submit/poll pattern:

Unique routing for special cases:

• 4K video → Kling V3 Pro
• Dialogue/lip-sync → Google Veo 3.1
• Multi-input (images + clips + audio) → Seedance 2.0
• Long-form video → LTXV-2

Text-to-Speech

3-tier routing with voice selection and quality gating:

Music Generation

3-tier routing with lyrics support and quality gating:

Self-Hosting & Private VPC

For organization-wide rollouts, high-throughput applications, or strict compliance frameworks, TemuClaude can be hosted in your own cloud infrastructure:

• Isolated Compute — Deploy the MoA pipeline inside your private AWS VPC, GCP project, or Azure subscription.
• Data Sovereignty — Ensure no transaction payloads ever leave your security perimeter by integrating with private frontier model instances (e.g., Azure OpenAI, AWS Bedrock).
• High Availability — Scalable Kubernetes templates and Serverless configurations optimized for zero-cold-start spikes.

Supported Environments

We provide officially supported deployment assets for the following platforms to commercial licensees: