Skip to main content
metalworks is the open-source core of Clique’s production pipeline. Clique runs on metalworks. This document explains the boundary and the principles behind it.

​
Why open-core

Two reasons. First, distribution: the easiest way to get a tool in front of the developers and agent-builders who would use it is to make the core free and embeddable. Second, it is a forcing function for clean architecture: extracting the pipeline behind versioned protocols, with no Supabase or auth singletons, makes it better software than it was as an internal monolith.

​
What is in metalworks

The pieces that are generally useful and carry no tenant-specific logic:
  • The data contract and the LLM / search / embedding / storage protocols.
  • The research vertical: brief to corpus to triage to clustered demand report.
  • The Reddit core: OAuth, search, metrics, inbox, subreddit intel, rate limiting, posting, and the deterministic compliance gate.
  • The discovery loop: filter to generate to gate.

​
What stays in Clique

Deliberately not open-sourced:
  • Account-backstory generation. The source had tooling that fabricated Reddit account histories to make accounts look aged and credible. That is coordinated inauthentic behavior and a Reddit ToS violation. It was excluded on purpose. The persona model here carries a background field that must be authentic; there is no path that invents one. See USAGE_POLICY.md.
  • The memory system. Clique’s episodic / structured / procedural memory is product surface. metalworks exposes a plain DiscoveryContext seam (voice guidelines, winning examples, pinned notes, personas) that a memory system renders into; the memory itself stays in Clique.
  • Multi-tenant auth, billing, plan caps, and spend caps. Tenant fields default to "local" so a library user never thinks about tenants.

​
Structural provenance

The research pipeline is built so that claims cannot be fabricated by construction, not by trust:
  • Cluster quotes are exact-matched against stored Reddit comments. A quote that does not match a real comment is dropped (no-quote-no-theme).
  • Web findings carry their source_url from the grounding tool’s citation metadata, never from model prose. Zero citations means the finding is dropped; URLs are never synthesized.
  • Counts (distinct authors, mention counts, must-address resolution) are computed from membership, never asserted by a model.

​
Internal vs external grounding

Web research follows the Inspect-AI split. When the chat model supports native grounding (Gemini’s google_search, Anthropic’s web_search), the adapter returns a GroundedResult with character-offset supports and the pipeline maps findings to citations by span overlap. When it does not, the pipeline falls back to an external SearchProvider (Exa, Tavily) plus a structured synthesis call that cites results by index. Either way, the provenance contract above holds.

​
API stability

Below 1.0, the stable surface is metalworks.contract and the MCP tool contracts. Everything else may change in any 0.x release. This lets the four form factors ship before the internal APIs are frozen.