Skip to main content
POST
/
api
/
v1
/
routing
/
resolve
curl -X POST https://api.manifest.build/api/v1/routing/resolve \
  -H "Authorization: Bearer mnfst_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {
        "role": "user",
        "content": "What is the capital of France?"
      }
    ]
  }'

{
  "tier": "simple",
  "model": "gemini-1.5-flash-8b",
  "provider": "google",
  "confidence": 0.9,
  "score": -0.3,
  "reason": "short_message"
}
The resolution endpoint analyzes incoming requests and determines the optimal tier and model based on request complexity, conversation context, and momentum.

​
Authentication

This endpoint uses Bearer token authentication with agent API keys (format: mnfst_*). 
curl -X POST https://api.manifest.build/api/v1/routing/resolve \
  -H "Authorization: Bearer mnfst_xxx" \
  -H "Content-Type: application/json" \
  -d @request.json

​
Request

​
messages
array
required
Array of conversation messages in OpenAI format
​
tools
array
Array of tool definitions (increases scoring complexity)
​
tool_choice
string | object
Tool choice strategy: auto, none, or specific tool object
​
max_tokens
number
Maximum tokens expected in output (influences tier scoring)
​
recentTiers
array
Array of recent tier assignments for momentum trackingValues: ["simple", "standard", "complex", "reasoning"]

​
Response

​
tier
string
Assigned tier: simple, standard, complex, or reasoning
​
model
string | null
Selected model name (null if no provider configured)
​
provider
string | null
Provider name (null if model not available)
​
confidence
number
Scoring confidence (0-1). Low confidence (less than 0.5) may fall back to standard
​
score
number
Raw complexity score from the scorer (can be negative for simple requests)
​
reason
string
Scoring reason:
  • scored: Normal scoring applied
  • formal_logic_override: Formal logic/math detected → reasoning tier
  • tool_detected: Tools present → minimum standard tier
  • large_context: More than 50k tokens → minimum complex tier
  • short_message: Less than 50 chars, no tools, no momentum → simple tier
  • momentum: Momentum from recent requests influenced tier
  • ambiguous: Low confidence → fallback to standard
  • heartbeat: Heartbeat request detected → simple tier

​
Request Example

curl -X POST https://api.manifest.build/api/v1/routing/resolve \
  -H "Authorization: Bearer mnfst_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {
        "role": "user",
        "content": "What is the capital of France?"
      }
    ]
  }'

​
Response Examples

{
  "tier": "simple",
  "model": "gemini-1.5-flash-8b",
  "provider": "google",
  "confidence": 0.9,
  "score": -0.3,
  "reason": "short_message"
}

​
Scoring Logic

Manifest’s scorer analyzes multiple dimensions:

​
Keyword Dimensions

  • Simple indicators: Greetings, basic questions
  • Reasoning indicators: Analysis, strategy, complex problem-solving
  • Code indicators: Implementation, debugging, refactoring
  • Writing indicators: Documentation, content creation
  • Math indicators: Calculations, equations, proofs

​
Structural Dimensions

  • Token count: Input length
  • Nested list depth: Task complexity
  • Conditional logic: Multi-step reasoning
  • Code-to-prose ratio: Code-heavy requests
  • Constraint density: Requirements and constraints
  • Expected output length: max_tokens parameter
  • Repetition requests: Revision/iteration keywords
  • Tool count: Number of available tools
  • Conversation depth: Number of turns

​
Overrides

  • Formal logic: Regex patterns for math/logic → reasoning tier
  • Short messages: Less than 50 chars → simple tier (unless momentum)
  • Tool presence: tools array with tool_choice != "none" → minimum standard tier
  • Large context: More than 50k estimated tokens → minimum complex tier

​
Momentum

The recentTiers array tracks recent tier assignments within a session. Momentum prevents rapid tier oscillation and keeps the agent in the appropriate tier for multi-turn tasks.
  • Momentum applies when recent tiers trend higher than scored tier
  • Bypassed for clearly simple messages (e.g., short messages matching simple indicators)
  • Helps maintain context during extended complex workflows

​
Tier Boundaries

Default scoring boundaries (configurable):
  • simple: score ≤ 0.1
  • standard: 0.1 < score ≤ 0.3
  • complex: 0.3 < score ≤ 0.45
  • reasoning: score > 0.45

​
Integration Notes

  • Most applications use the proxy endpoint (/v1/chat/completions) which calls resolve internally
  • Use this endpoint directly only if you need tier information before making the actual LLM request
  • The proxy applies additional optimizations:
    • Filters out system/developer roles before scoring (prevents inflated scores)
    • Takes only last 10 messages for scoring
    • Detects OpenClaw heartbeats and routes to simple tier

Build docs developers (and LLMs) love

Get started for freeTalk to us