An open-source agent flow for front-end chip design

VeriPower is an open-source agent flow that takes a natural-language spec all the way to front-end signoff on commercial EDA tools. A deterministic engine sits underneath, recording every action in an append-only log. All pipeline status is derived from that log on demand, never stored as a flag or snapshot. The agent can iterate on its own, but every LLM-authored oracle needs a human sign-off.

Ships as a plugin for Claude Code, opencode, and DeepSeek Harness.

How it works

A deterministic engine owns the facts. Agents and humans are proposers. The engine keeps an append-only record of every action and conclusion, and whether a stage is done, stale, or failed gets computed from that record each time you ask. An orchestrator queries the engine for one action, carries it out, queries again. No state carried between queries.

VeriPower architecture

Every verification conclusion is fingerprinted against the content it consumed and produced. Edit something upstream and the downstream conclusions go invalid on the next query. Design and verification both start from the spec but then diverge, so the reference model is derived from the spec, not from the implementation. EDA tool verdicts are authoritative. LLM-authored oracles are good enough for iteration but need human endorsement for signoff, and that endorsement lapses if the oracle content changes.

Pipeline

Eight stages, spec through power analysis. The dependency graph falls out of each rule's artifact declarations.

Pipeline dependency graph

Reference implementation wraps Synopsys tools (SpyGlass, Design Compiler, PrimeTime, VCS+UVM). Each stage is a self-contained skill, so you can swap one (Verilator for simulation, Yosys for synthesis) without touching the rest.

Results

Three front-end design tasks, bare Claude Code vs. Claude Code + VeriPower. Same LLM, same spec, same EDA tools. Paper forthcoming.

Benchmark	Scale	Baseline	+ VeriPower
gateGPT — fixed-point GPT inference	346K gates	3 of 4 coverage metrics below 90%	pass
FSA — FlashAttention accelerator	32K gates	2 unresolved CDC violations	pass
Coral-NPU — RISC-V ML accelerator	3.4M gates	6/19 tests	19/19

Bare agents write correct RTL but can't catch their own blind spots when testing it.

Quickstart

Claude Code

claude plugin marketplace add chipweaver/veripower
claude plugin install veripower@chipweaver

Or point at a working copy: claude --plugin-dir /path/to/veripower.

opencode — add the plugin to ~/.config/opencode/opencode.json, or to a project-level
opencode.json:

{ "plugin": ["veripower@git+https://github.com/chipweaver/veripower.git"] }

Stage dispatch runs subagents in the background, which opencode gates behind an environment
variable, so start it with:

OPENCODE_EXPERIMENTAL_BACKGROUND_SUBAGENTS=true \
OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX=131072 opencode

Without the second flag, opencode (as of 1.18.x) caps every completion at 32,000 tokens
regardless of the model's declared limit. Subagents authoring whole-module RTL think past
that cap and die silently. 131072 matches the GLM-5.x declared limit, and models declaring
less keep their own.

DeepSeek Harness — install into the profile you run:

dsh plugin --profile web add "veripower@git+https://github.com/chipweaver/veripower.git"

Run the web profile (dsh web), not the one-shot headless profile.

Ask it to list its skills — the twelve VeriPower ones confirm the install.

Run the brainstorm skill to settle requirements first, then tell the agent:

Run the design flow for {module_dir}

Full walkthrough in the user manual (中文).

Requirements. Python 3.10+, jsonschema >= 4.18, referencing, PyYAML. Synopsys EDA tools, swappable per skill.

Citation

Paper forthcoming.

Status: alpha (v0.2.0). MIT License. Contributing. Issues.