Skip to main content

Documentation Index

Fetch the complete documentation index at: https://manifest.build/docs/llms.txt

Use this file to discover all available pages before exploring further.

​
Agent

A configured client that sends requests through Manifest. Each agent has its own API key (mnfst_...), its own routing rules, and its own usage page. An agent typically corresponds to one tool or workflow (your IDE plugin, a Slack bot, a scheduled job), not one user.

​
Auth type

The credential category Manifest uses to talk to a provider:
  • api_key — classic per-token API key (most providers)
  • subscription — OAuth or token tied to a paid plan (ChatGPT Plus, Claude Max, GLM Coding, etc.)
  • local — no credential, server runs on your machine (Ollama, LM Studio, llama.cpp)
Auth type is recorded on every request and shows up in the dashboard’s distribution chart.

​
Confidence

A score between 0 and 1 indicating how clearly a request fits its assigned tier. Returned in the X-Manifest-Confidence response header.

​
Fallback

The retry mechanism that kicks in when the primary model fails. Manifest tries the next model in the tier’s fallback list, then the next, until one succeeds or the list is exhausted. See Fallback for triggers and config.

​
Fallback chain

The ordered list of models tried for a single tier or specificity, primary first. Up to 5 models. When all of them fail, Manifest returns HTTP 424 with X-Manifest-Fallback-Exhausted: true.

​
Manifest/auto

The model ID clients send to opt into routing. Any other model ID is forwarded as-is to the matching provider. manifest/auto is the only string that triggers tier scoring and model selection.

​
Momentum

The carry-over rule that prevents a short follow-up (“yes”, “do it”) from dropping to a cheaper tier. See Routing → Session momentum.

​
Provider

An upstream LLM service Manifest can route to. There are four kinds:
  • API key — pay-per-token (OpenAI, Anthropic, Google, …)
  • Subscription — flat-rate plan (ChatGPT Plus, Claude Max, …)
  • Local — runs on your hardware (Ollama, LM Studio, llama.cpp)
  • Custom — any OpenAI- or Anthropic-compatible HTTP endpoint

​
Routing

The scoring + selection process that picks which model handles a request. Two axes drive it: tier (how complex) and specificity (what kind of task). Both happen in under 2 ms with no external calls.

​
Scoring

23 weighted dimensions (14 keyword-based, 5 structural, 4 contextual) that produce a single complexity score. Threshold boundaries map the score to a tier. Same pipeline feeds the specificity detector.

​
Specificity

The task category Manifest detects from keyword patterns and tool names. Nine categories: coding, web_browsing, data_analysis, image_generation, video_generation, social_media, email_management, calendar_management, trading. When the match crosses the threshold, you can pin a category-specific model that overrides the per-tier choice. See Routing → Task-specific.

​
Tier

The complexity bucket a request falls into: simple, standard, complex, or reasoning. Each tier maps to a primary model plus a fallback chain. Tier is the main lever Manifest pulls to balance cost against capability. See Routing → Complexity.

​
Tier override

Any signal that forces a minimum tier regardless of the computed score. See Routing → Tier overrides.