sources

Overview ¶

Package sources — recipes for AI source / gateway integrations.

First entry: bifrost-template. maximhq/bifrost (https://github.com/maximhq/bifrost, Apache-2.0) is a Go-native AI gateway with unified failover, semantic caching, and budget governance. The portal layer registers a `bifrost` driver (internal/portal/bifrost.go) that surfaces in `clawtool portal list` as `bifrost (deferred)` until phase 2 lands the bifrost/core dependency behind the `clawtool_bifrost` build tag.

This recipe ships the YAML config template the driver will read once phase 2 lands. Dropping it today is safe — nothing reads the file yet, but operators get a head start on writing their provider chain + cache settings.

guardians-stub recipe — drops a pre_send rule template that exercises clawtool's `guardians_check(plan_arg)` predicate.

metareflection/guardians (https://github.com/metareflection/guardians, MIT) implements Erik Meijer's "Guardians of the Agents: Formal Verification of AI Workflows" (CACM, January 2026): taint-tracking + Z3-SAT verification of an agent's drafted plan before SendMessage dispatches. Phase 1 (this commit) ships:

• the `guardians_check(plan_arg)` predicate in internal/rules, wired as an always-true stub so the surface contract exists;
• this recipe, which drops a sample rules.toml fragment under .clawtool/rules/ that operators can fold into their main rules file or load directly.

Phase 2 will flip the predicate to invoke the Z3-SAT engine, gated behind a `clawtool_guardians` build tag so the Z3 cgo dependency stays opt-in. The rule shape doesn't change between phases — the same `guardians_check("plan")` call routes through the stub today and the real engine tomorrow.

Package sources hosts recipes that wire up MCP source servers whose configuration ships as a templated file in the repo. The canonical example is googleapis/mcp-toolbox (Apache-2.0): a DB MCP server that reads a tools.yaml describing one or more database sources + parameterised SQL tools.

Recipes here ship config only — the operator installs the upstream binary themselves. Category routes through CategoryRuntime because DB integration is dev-environment scaffolding.