Empryo
A graph-aware coding agent that changes code through AST symbols instead of text patches.
What does this agent do, and when should you use it?
Empryo is a locally run AI coding agent formerly known as SoulForge. On launch, it uses tree-sitter to parse a repository into a live graph of symbols, imports, and call sites, ranked with PageRank and Git co-change data. It performs symbol-level AST edits rather than find-and-replace, with atomic batches and rollback. The product is offered through a native desktop app, a full terminal UI, and a headless CLI for scripts and CI, all using the same code genome. This repository serves as Empryo’s public issues and discussions home while retaining archived SoulForge source; active new-feature development has moved to Empryo.
After you run empryo in a project, it parses the repository, indexes symbols, imports, and call sites, and builds a queryable code graph. Before editing, it can surface import relationships, historical co-changes, and blast radius; it then changes code through 65+ symbol-level AST operations and structural edits across 30+ languages. Batches are atomic with rollback, and typechecking is described as the gate for edits. The product also documents parallel explore/edit agents with a shared I/O cache, ten routable roles, MCP server support, language-server support, and Git checkpoints for each prompt.
- A developer refactoring a multi-module repository who wants to inspect call sites and likely blast radius before changing a symbol.
- An engineer working primarily in a terminal who wants to run `empryo` directly inside the target project.
- A team adding a headless coding-agent step to a script or CI workflow while using the same code graph as the interactive interfaces.
- A developer with several workspaces who wants one tab routed to a low-cost scouting model and another to a stronger coding model.
- A team handling private repositories locally and choosing either its own provider keys or a local Ollama/LM Studio model setup.
What are this agent's strengths and limitations?
- Navigates through a live graph of symbols, imports, and call sites, with PageRank and Git co-change ranking instead of relying solely on text search.
- Offers 65+ symbol-level AST operations, atomic batching, and rollback for structural changes rather than fragile string replacement.
- Covers native desktop, terminal UI, and headless CLI workflows with one shared code genome.
- Documents routable multi-model roles, 22 providers, local Ollama/LM Studio operation, and OpenAI-compatible endpoints.
- Cloud-model use requires bringing your own provider credentials; local no-key use still requires an Ollama or LM Studio environment.
- The repository says active feature development has moved to Empryo, while this repository is principally for issues, discussions, and archived SoulForge source, so it does not expose the current Empryo implementation for review.
- The documented installation paths include remotely fetched shell or PowerShell scripts, which may conflict with organizational installation or supply-chain policies.
- Advanced language-server functionality depends on language servers managed through Mason, so practical coverage depends on the relevant server and local setup.
How do you install or deploy this agent?
On macOS/Linux: curl -fsSL https://empryo.com/install.sh | bash. On Windows PowerShell: irm https://empryo.com/install.ps1 | iex. Package-manager options are brew install proxysoul/tap/empryo and winget install ProxySoul.Empryo. Configure Anthropic with empryo --set-key anthropic sk-ant-...; alternatively, run with local Ollama without a key.
How do you use this agent?
Change to the project and start the terminal UI with cd your-project followed by empryo. The documented download page also provides the desktop app and prebuilt binaries. Role routing can be configured globally, per project, or per tab; the README also documents MCP servers and language-server capabilities.
How does this agent compare with similar options?
The README reports benchmark comparisons with pi. Across two stated rounds, Empryo reports lower cost and shorter wall-clock time; the second round uses real bugs from hono, zod, and ky. The comparison is limited to the models, repositories, and tasks described there.