automoat
Planner How it works Examples Measurement Evidence
Local AI · Useful by design

Put your local tokens to useful work.

Choose the memory you have—or want. Tell Automoat what repeats in your work. It will build a bounded local-agent plan, or suggest a useful starting point if you are beginning with spare compute and no moat idea yet.

Build my local plan See how Automoat decides
What is a moat here?
One recurring job, private context generic models do not have, an objective verifier, and recorded feedback that makes the job improve over time.
Build your operating plan Step 1 of 4
What should your local agent run on?

Start with the machine decision. The plan will keep model size provisional until a real local benchmark exists.

Your starting point
Hardware type
Memory available or desired

Choose the memory available to inference. Automoat will benchmark the actual machine before promising throughput.

1 Calibrate the machine

Memory sets a safe model band. A short local receipt measures real speed, token use, and task quality before the budget expands.

2 Choose a compounding job

Automoat joins recurring work with private context, an artifact, and a verifier. If any piece is missing, it is not a moat yet.

3 Spend, verify, promote

Useful work gets a bounded queue. Verification and corrections earn future token budget; unverified activity stops early.

Useful starting points

A moat can start as one job that gets measurably better.

Local token yield labTest model, prompt, context, and task combinations; promote the route with the best held-out value per token.
Nightly codebase caretakerTurn failing checks and explicit backlog signals into one tested patch or an evidence-backed investigation.
Workflow exception finderReview recurring records, rank exceptions, and propose the next bounded action against known outcomes.
Build and runtime evidence

Watch the proof pipeline work.

Render runs Codex, GitHub records the work, and Vercel exposes the live stream. This is inspectable development and runtime evidence for Automoat, not the product identity or the local inference path. It shows how bounded improvements are checked and recorded.

Runtime Render worker
Trail commit + push
Visibility health-checked log relay
render codex relay connecting
$ codex loop --autonomous --publish-cockpit --push main 
connecting to Render Codex worker...
Loop

The agent acts on new evidence, runs checks for material work, and records no change when the evidence is unchanged.

Memory

Each correction, failure, passing contract, and operator decision becomes future context.

Artifacts

The work produces files, evals, contracts, logs, and handoffs instead of disappearing into chat.

Current initiative · Local Run Receipt

One receipt for speed, token cost, privacy, and task quality.

The first local-inference harness runs an immutable JSONL task pack against an OpenAI-compatible model endpoint. It permits loopback endpoints by default, refuses remote inference without an explicit opt-in, and records measurements without copying raw business inputs or model outputs into the receipt.

Measured, not marketed Prompt, completion, and total tokens sit beside wall time and end-to-end output tokens per second. An optional compute-hour rate yields effective cost per million tokens and per strict exact match.
Reproducible boundary The task-pack digest, model identity, runtime, hardware description, and active optimization travel with the result. A changed task pack cannot inherit an old comparison.
Moat comparison A candidate run can compare with a baseline only when both receipts bind the same task digest. The useful question is quality per token on the work your business owns.
Run the local harness Read the DFlash2 example

DFlash2 is one example of the moving local-inference frontier, not a bundled Automoat dependency and not a universal speed claim. The harness keeps speculative decoding, quantization, caching, and future runtime techniques replaceable and proves each one on the actual hardware and task pack.

Local Run Receipt capability · Private Workload Fit

Know if this run met your work's limits.

Set the task-pack digest and your quality, time, and estimated-cost limits first. Automoat reads an existing content-free run receipt and returns a short decision card: fit, not fit, or unmeasured for that observed sequential task pack. You see the exact reason and the next bounded measurement, not another runtime leaderboard.

What was observed Strict task matches, measured sequential wall time, operator-rate estimated cost, and whether the request addressed a loopback endpoint.
What stays unknown A sequential receipt cannot establish parallel capacity, peak memory, cache reuse, or verified no egress. Loopback address scope is not a listener or telemetry audit.
Your decision remains yours The card never starts a model or agent, chooses a runtime, downloads weights, or grants authority. Any runtime speed claim still needs its own eligible-machine proof.
Read the policy and CLI contract

A fit result applies only to the supplied receipt and frozen policy. It is not a general model benchmark, a privacy guarantee, or an autonomous work permission.

Released capability · Moat Builder

Whole-Record Check

One current application of Automoat's core loop: turn bounded workflow evidence into an inspectable recommendation, then spend human attention only on material disagreement.

A recommendation that agrees with its immutable case snapshot earns a Coverage Receipt. Missing actions or material evidence become Evidence Conflict cards with stable source IDs and correction-ledger context, so operators review the exception instead of rereading every record.

Dallas validation set 30 versioned Dallas scaffold cases are stored as bounded, content-addressed snapshots.
Quiet agreement 20 complete candidates produce Coverage Receipts with their action and source coverage.
Visible exceptions Exactly 10 planted retrieval omissions produce 10 Evidence Conflict cards and no unexpected conflicts.
Open the 30-case evidence Inspect report JSON

Dallas validation boundary: this is a deterministic regression harness over versioned scaffold records, not a production model-accuracy benchmark. Dallas is not the full Automoat product.

Release Notes

What changed

2026-09-19

Added Private Workload Fit beneath Local Run Receipt: a content-free decision card that checks one observed sequential task pack against a frozen human policy. It reports fit, not fit, or unmeasured with the exact reason and next measurement; it does not infer parallel capacity, memory, cache reuse, or verified no egress.

2026-08-30

Turned the landing page into the product: a four-step local-token planner that matches memory, idle capacity, recurring work, privacy, autonomy, and verification. It can shape an existing moat idea or suggest one, returns a downloadable operating plan, and falls back to a private rule-based starter when the hosted planner is unavailable.

2026-08-27

Recentered Automoat on its two connected systems: measurable local AI on consumer hardware and a harness that compounds proprietary work into a proved moat. Added the Local Run Receipt contract so token economics, the device boundary, runtime provenance, and task quality stay in the same artifact; Whole-Record Check remains a released Moat Builder capability rather than the product identity.

2026-08-19

Made cockpit health fail honestly. The relay badge now follows cockpit_ok, cockpit_status, and the health label instead of calling a stale or degraded snapshot live; the pulsing indicator is reserved for a confirmed live relay.

2026-08-19

Restored Automoat's durable product hierarchy. The root now leads with the local-first job of discovering, defining, proving, and operationalizing a moat; business-first and dataset-first entry points come next. Whole-Record Check is a released current initiative, with Dallas kept explicitly as its bounded validation case rather than the product identity.

2026-08-19

Released Automoat Whole-Record Check: a backend-neutral verification pass over 30 immutable Dallas case snapshots. The deterministic validation plants exactly 10 retrieval omissions, detects all ten as Evidence Conflict cards with stable source IDs, and issues 20 Coverage Receipts on agreement. Results carry correction-ledger-compatible queue metadata without writing to the operator ledger.