The pipeline

One install intercepts each agent request at a single point and runs it through five stages, always in the same order. Every product you use is one of these stages, and every stage writes what it learns to the one signed knowledge network, Weave.

The five stages

request ─→ security ─→ cost ─→ quality ─→ observability ─→ memory ─→ response
              Guard      Cap      Bench      Radar          Mind
              (all on your device, one interception point)

• Security (Guard) - scan for secrets, injection, and risky commands first, before anything else runs. Fail-closed: on doubt, block.
• Cost (Cap) - preserve prefix caches byte-exact, prune stale context, and measure the cache-write premium against how often it is read back.
• Quality (Bench) - measure model performance on your repo so routing decisions are grounded in your tasks, not a generic benchmark.
• Observability (Radar) - record a cross-agent timeline so drift is caught early and any session can be rewound.
• Memory (Mind) - write durable facts to Weave with provenance so the next session starts informed.

Why the order is fixed

Security runs first so nothing dangerous reaches the later stages. Cost runs before quality so routing decisions account for price. Memory runs last so it captures the outcome of everything upstream. The order is an invariant - it is not configurable per request, because a reordered pipeline would let an unsafe call slip past Guard.

One spine, many products

Each stage is a product you can adopt independently, but they share one event spine and one Weave. That is what makes the whole greater than the parts: Cap's routing uses Bench's measurements; Radar's timeline references the same facts Mind stores. There is never a second, competing interception layer - see the One-Layer rule.