Sources and Integrations

Sources & Integrations

UASL integrates external skill sources and optional tools by reference, not by vendoring. The canonical corpus stays in skills/; external roots are indexed read-only, and optional services are reached through router pointers that keep their auth/state local.

Source docs: docs/SOURCE_INTEGRATIONS.md · docs/THIRD_PARTY_SOURCE_REPOS.md

---

Non-redundancy rule

The canonical corpus is single-sourced in skills/. External sources are indexed read-only; nothing is copied into the repo.

This means the router can see skills that live in other tools' directories without the repo carrying a second copy that could drift. Secrets and runtime state always stay in the user's own home directory, never in the repo.

---

Third-party skill sources (indexed read-only)

Source	Local path (typical)	Notes
GStack	~/.gstack/gstack	Engineering skill stack, maintained separately. Router sub-sources: gstack-gbrain, gstack-codex, gstack-openclaw.
GBrain	~/gbrain	Personal knowledge brain; indexed read-only as a memory/source pointer.

GStack's router sources map to:

• gstack-gbrain → ~/.gstack/gstack/.gbrain/skills
• gstack-codex → ~/.gstack/gstack/.agents/skills
• gstack-openclaw → ~/.gstack/gstack/openclaw/skills

---

Optional services & tool wrappers (router pointers only)

These are reached through a router skill/pointer; their credentials and artifacts stay in the user's account or home directory.

Integration	What it adds	Local state
MemPalace	Durable memory store	~/.mempalace/palace
NotebookLM	Notebook research source	Auth/artifacts in the user's Google account
x-cli	X (Twitter) API workflows	~/.xrc
Instagram CLI	Instagram workflows	~/.instagram-cli
Crawl4AI	Local, LLM-ready crawling	~/.crawl4ai
Firecrawl	Hosted web-data API/CLI	Keys kept local
Context Mode	Context-window protection (not durable memory)	Host tool/wrapper
Lightpanda	On-demand browser automation	Host tool/wrapper
OpenRouter	Optional OpenAI-compatible model fallback	OPENROUTER_API_KEY env var

Note

Web search is host-native by default — no provider keys live in the repo. Lightpanda is used only for controlled, on-demand retrieval.

---

The local runtime (Universal AI Stack)

On a full install, ai-setup/ materializes a local runtime that routes model calls across providers with health checks and a guarded local fallback:

• Router endpoint: http://127.0.0.1:18100/v1, model auto-coding.
• Default model failover order: GPT → Kimi (Moonshot, HTTP fallback) → Claude Opus → local Qwen (llama.cpp).
• Config (repo-owned source of truth): ai-setup/runtime/config/model-registry.json, routing-policy.json (240s global timeout, ~10-min circuit breaker), source-integrations.json.

See Installation & Setup for how the stack is laid out and which files are repo-owned vs machine-owned.

---

How a source becomes routable

1. A pointer/config entry registers the external root in source-integrations.json.
2. The router indexes that root read-only alongside the canonical corpus.
3. Prompts route to skills from any indexed root — without duplicating them into the repo.

This is the same decoupling principle as the rest of the architecture: the router resolves locations; it never owns the content.