Private beta · v0.4

Pipelines for resilient AI agents.

ClawPipe orchestrates the full agent loop — task graph, deterministic execution, durable memory, verification — so your agents ship work, not chat transcripts.

Request access → See how it works

clawpipe ▸ run property-watcher

// Define once. Re-run forever. const pipe = clawpipe.define({ id: "property-watcher", steps: [ { id: "fetch", fn: fetchListing, retries: 3 }, { id: "redact", fn: redact, deps: ["fetch"] }, { id: "verify", fn: verifySchema, deps: ["redact"] }, { id: "notify", fn: sendOwnerWA, deps: ["verify"], idempotent: true }, ], }); await pipe.run({ ref: "68899" });

● fetch · 88ms ● redact · 4ms (2 fields) ● verify · 12ms ● notify · 143ms

// What ClawPipe is

Six primitives that turn a chatbot into an agent.

Open-ended reasoning + deterministic workflows + verifier layer + durable memory + idempotent effects + trust-graded autonomy. Each one a separate concern, all wired together.

Task graph, not transcripts

Goals decompose into typed tasks with dependencies. Retries, fan-out, sub-tasks, approvals — first-class primitives, not prompt tricks.

Verifier layer

Planner → executor → verifier → optional human approval. Nothing ships until checks prove it shipped.

Durable memory

BM25 + vector + RRF fusion across artifacts, decisions, runs. Inspectable. Editable. Versioned. Dedup-aware.

Idempotent effects

HTTP, WhatsApp, email, payments. Idempotency keys, retry policies, compensating actions. No double-sends.

Durable waits

Pause for minutes, hours, or days without losing run state. Resume exactly where it left off — no manual stitching.

Trust-graded autonomy

Skills earn autonomy through measured production success. Approvals stay where blast radius is real.

// How it runs

Every run, the same five rails.

Goal in. Artifact out. Every step traced, scored, and re-runnable from any phase.

Goal

Plan

Execute

Verify

Ship

      checkpoints · resume from any phase
      artifacts · every step writes a file
      trace · full trajectory, not just outcome
      eval · pass-rate tracked per skill
    

// Stack

Adapters over lock-in.

Bring your own model, vector store, queue, and side effects. ClawPipe normalizes the contract — you keep ownership.

Models

Routed by complexity, latency, and budget. Fall back to local when external is down.

Claude
GPT-5
Gemini
Ollama
vLLM

Memory

SQLite + FTS5 + vector embeddings out of the box. Swap to pgvector, Qdrant, or LanceDB at scale.

SQLite FTS5
MiniLM-L6 384d
pgvector
Qdrant

Effects

HTTP, WhatsApp, email, calendar, Stripe, GitHub. Idempotency keys + retries + compensations baked in.

WhatsApp
SMTP
Stripe
GitHub
HTTP

Runtime

Self-hosted Node.js + SQLite. Add Postgres + Redis when you need horizontal scale. No vendor lock-in.

Node 22
SQLite
Postgres
Redis
PM2