Empryo

Empryo

_{previously SoulForge}

The AI coding agent that edits symbols, not strings.

Empryo reads your repository as a living system: it sequences a genome — a live, ranked dependency graph of every symbol — and edits through the AST, never find-and-replace.

Website · Download · Benchmarks · Changelog · Discussions · Discord

SoulForge is now Empryo — the same symbol-level, graph-powered agent, rebuilt with a desktop app, a faster engine, and a composable core. This repository is Empryo's public home for issues and discussions.

# macOS / Linux
curl -fsSL https://empryo.com/install.sh | bash

# Windows (PowerShell)
irm https://empryo.com/install.ps1 | iex

brew install proxysoul/tap/empryo
winget install ProxySoul.Empryo

empryo --set-key anthropic sk-ant-...   # or run locally with Ollama — no key required
cd your-project
empryo

Desktop app and prebuilt binaries: empryo.com/download. Runs on macOS, Linux, and Windows.

Most coding agents grep, read whole files, and patch strings — they never know what depends on the code they just changed. Empryo builds understanding before it mutates anything:

• It maps before it reads. On launch, tree-sitter parses your repo into a live graph — every symbol, import, and call site, ranked by PageRank and git co-change. Graph queries answer in milliseconds and cost zero LLM tokens.
• It knows the blast radius. Before an edit, the agent sees what imports a file, what historically changes with it, and how far a change ripples — "what breaks if I touch this?" is answered before the first keystroke.
• It edits through the AST. 65+ symbol-level operations, atomic batches with all-or-nothing rollback, structural edits across 30+ languages, and a typecheck as the gate. Nothing breaks on whitespace.
• It treats tokens as spend. The graph does the navigation models usually burn context on — fewer reads, fewer steps, smaller bills at any scale.

Empryo isn't one model in a loop — it's a crew, and you assign the seats. Every role is a routable slot that takes any model from any of the 22 providers:

• Per tab. Each workspace tab carries its own routing — a frontier model writing code in one tab, a fast cheap one triaging issues in the next, a local model on a private repo in a third.
• Per config. Set defaults globally or per project; override any slot from the tab. Cheap models scout, strong models write, reviewers judge with clean context.
• Custom agents. Define your own agents — a prompt, a model, a tool policy — and dispatch them alongside the built-ins. Mix and match providers freely inside a single run.
• Cache-aware by design. Routing keeps prompt-cache prefixes stable — sub-agents inherit their parent's cache line, so repeated context bills at cache-read rates instead of full price.
• Costs, itemized. Live spend tracking per model, per sub-agent, per tab, per session, per day — you always know where the tokens went.

Head-to-head against pi — same models, same repositories, same tasks. The graph does more with less:

Round 2 used real bugs from merged PRs (post-training-cutoff, history scrubbed, regression tests injected after each run). Full methodology and transcripts: empryo.com/benchmarks · reproduce at proxysoul/pi-vs-empryo-bench.

Empryo runs entirely on your machine. Bring your own key — Anthropic, OpenAI, Google, Groq, DeepSeek, Bedrock, and 16 more, or any OpenAI-compatible endpoint — or run fully local with Ollama / LM Studio. No proxy in the middle, no code leaving your machine, no per-seat fee. Free to use.

SoulForge remains available to download and install, and continues to receive fixes for bugs and critical issues. New features and active development have moved to Empryo.

brew tap proxysoul/tap && brew install soulforge
# or
bun install -g @proxysoul/soulforge

• Issues and Discussions — the home for Empryo bug reports, questions, and ideas.
• The SoulForge source remains archived here under its existing license (see LICENSE).

_{One API, 200+ models, up to 30% off frontier. Wired in as the llmgateway provider.}

_{Sponsor · PayPal · All backers}