Skip to main content
A fork of Zed · master agent JouleClaw

Deterministic-first coding.
The model is the last resort.

Agent Zed is a fork of Zed. Its master agent, JouleClaw, resolves most code through lookup, search, and templates — local, free, deterministic. When a model is genuinely needed, JouleClaw's rigid harness elevates whatever you point it at: a free-tier endpoint, a paid API, or an openweight model on your own hardware. The harness does the work, so the model doesn't have to be the biggest — and where others take a random walk, it stays deterministic.

LUT
deterministic
Search
deterministic
Template
deterministic
Model
last resort
The cascade

Deterministic-first

The cheapest method that can close the change, wins.

JouleClaw runs every request down a cascade and stops at the first tier that can resolve it. Lookup, search, and templates — deterministic, hardware-optimized, near-free — take precedence. Generation is the exception, not the rule, and a model — any model you bring — is the very last thing tried.

Lookup
LUT / cache det

The change was resolved before. A content-addressed lookup returns it instantly, for picojoules — exact, reproducible, free. The first thing JouleClaw tries.

Search
retrieval det

Answerable from a search over your tree and prior changes — the nearest known solution, surfaced without generating anything new.

Template
deterministic rule det

A formatter, a codemod, a rename, a scaffold — a transformation with one correct answer, applied by rule. No model involved.

Local model
on-device inference gen

When generation is genuinely needed, a local model runs first — on your hardware, metered in joules, nothing leaving the machine.

Any model
last resort · your key gen

When nothing deterministic can close, JouleClaw invokes a model — free-tier (OpenRouter, Lightning AI), paid API, or openweight, set with your own key. Its rigid harness elevates whichever you bring. On most changes, it's never reached.

The master agent

One rigid harness that elevates every model.

The agent in charge isn't a model — it's JouleClaw. It owns the loop and wraps every model call in a rigid, effective harness: structure, retrieval, templates, and verification do the heavy lifting, so the model fills a tight, well-formed gap instead of facing a blank page. That's the meat — a free or small model inside this harness outperforms a frontier model left to wander. Which model you bring is your call.

Other tools — a big model, left to wander
large model
Nondeterministic, expensive, hard to reproduce.
JouleClaw harness — model fills a tight gap
look search tmpl model verify one rigid path
Deterministic, reproducible — any model works, even a small one.

Bring any model

Free-tier endpoints (OpenRouter, Lightning AI), paid APIs, or openweight models on your own hardware. Set your own keys, swap any time — JouleClaw is aware of all of them and locks you to none.

The harness elevates it

JouleClaw's rigid scaffolding — lookup, search, template, structured tool use, verification — is what produces quality. The model only fills the gap the harness leaves, so a modest or free model becomes genuinely effective.

Cheaper, and it shows

Because the deterministic tiers close most work and the harness elevates whatever runs, the model can be the cheapest one that fits. The receipt shows exactly when a model was — and wasn't — needed.

Energy tracking

Every change in real joules — on every hardware ecosystem.

Deterministic tiers cost almost nothing; generation costs real energy. JouleClaw meters the work from the host's own counter wherever the hardware exposes one, so the status line — ⚡ N J this session — shows you precisely when you needed the expensive path and when you didn't. The same probe, the same way, across the ecosystems developers use.

Energy per close, by tier
LUT Search Tmpl Local Model energy / close →
Illustrative — the deterministic tiers cost almost nothing, so most sessions barely move the meter.
Hardware Counter Provenance Reading
Apple Silicon IOReport Energy Model ModelBased 25.5 W / 74 J
Windows + RTX 2060 NVML HwShunt 9.4 W / 18.8 J
Linux x86 + Tesla T4 NVML HwShunt 71.8 W / 148 J
Linux x86 Xeon RAPL powercap HwShunt 77.1 W / 154 J

Verified on real silicon. Each reading declares its provenance — HwShunt for a hardware counter, ModelBased for a vendor energy model — so the number is honest about how it was known.

Code traceability

Every edit is an object you can name, sign, and account for.

JouleClaw captures each change as a content-addressed, signed coding operation with a receipt — not a disposable diff. The full provenance of your code: what changed, on what base, at which tier, for how much energy.

JCR-1 RECEIPT det-CBOR · COSE_Sign1
op
b3:Ynpl53wU4pwfokr8h2c…
prior
b3:8fK2q9Lm4nRtZ0xV…
tier
L1 · Template
energy
0.004 J · HwShunt
sig
ed25519 ✓ verified

Content-addressed

Every change JouleClaw produces is named by its BLAKE3 hash (b3:…) over a canonical encoding. The same change always has the same id — a stable fingerprint for every edit.

Signed & attributed

Each operation is an Ed25519-signed transition bound to the prior state of the tree. Who changed what, on what base, by what tier — provable, not assumed.

Receipted

Each change carries a receipt: the tier that produced it and the joules it cost, with honest provenance. You can see whether a lookup or a frontier model did the work.

Reproducible by construction

The deterministic tiers — lookup, search, template — give the same output for the same input, every time. Most of your history is replayable, not a random walk.

Get it

Install Agent Zed.

A fork of Zed, in the open — Apache-2.0 lineage, pinned to the Transaction Science pillars by git revision. Install it, connect a model key, and let JouleClaw reach for the deterministic answer first.

macOS · Homebrew
brew install --cask transaction-science/tap/agent-zed

Apple Silicon & Intel. Or the .dmg from Releases.

Linux · apt
sudo apt install agent-zed

x86_64 & arm64, after adding the apt repo. Or .deb / .tar.gz.

Windows · Scoop
scoop install agent-zed

x86_64 & arm64, after scoop bucket add transaction-science …. Or the .exe installer / portable .zip from Releases.

# or build from source
git clone https://github.com/Transaction-Science/agent-zed
cargo build --release --bin agent-zed
# point it at any model — free-tier, paid, or openweight — with your own key