The agent that treats code as code

~5 tokens to change a function return type (AST edit vs ~100 lines of oldString)

~$0 average compaction cost (V2 extracts as you go)

34 → 5 messages after compaction with 0 LLM tokens spent

Every other AI coding tool treats your codebase as text. It greps, it pastes 500-line files into context, it builds oldString/newString blobs and prays the whitespace matches. Half the turn is orientation. The other half is string-matching roulette.

SoulForge treats code as code. On startup it parses your project into a live Soul Map — every file, symbol, import edge — ranked by PageRank and git co-change. Forge opens every turn oriented: it reads single symbols by name (not whole files), edits TS/JS through the AST (symbol kind + name, 65+ ops, zero text matching), and rewrites the rest with line-anchored edits that never drift.

What that means in practice:

• ast_edit changes a return type with value: "Promise<User>" — no oldString, no whitespace failures, no line-offset math.
• Reads pull one function by name, not an 800-line dump. Context stays lean.
• V2 compaction serializes structured state as the conversation happens — when context fills, it compacts for free.
• Sub-agents share a read cache, so 3 parallel explorers don't re-open the same file 3 times.

Same work, a fraction of the tokens, a fraction of the seconds.

Not your average CLI

🧠 Live Soul Map SQLite graph of every file, symbol, and import — PageRank-ranked, git-co-change-aware, personalized per turn. Renders into the system prompt with blast radius tags so the agent knows which edits ripple. Learn more	🔪 Surgical reads across 33 languages Read a single function by name. A 500-line file becomes a 20-line extraction. TypeScript, Python, Rust, Go, Java, Ruby, C/C++, Swift, Kotlin, Elixir, Zig, Solidity, and more. Learn more
🤖 Parallel agents with shared cache Forge dispatches explore, code, and web-search agents in parallel. Files one reads are cached for the others — 3 agents don't re-read 3x. Real-time findings propagate between them. Learn more	💰 Free compaction V2 compaction tracks structured state as the conversation happens — files touched, decisions, failures, tool results. When context fills up, serialization is instant and typically costs zero LLM tokens. Learn more
🎯 One call, complete job rename_symbol runs LSP rename, verifies zero dangling refs, reports back. move_symbol moves a symbol and updates every import across TS/JS, Python, and Rust. project auto-detects your toolchain across 23 ecosystems. Learn more	🔬 AST-native editing ast_edit addresses TS/JS symbols by name, not text. 65+ operations. Changing a return type costs ~5 tokens instead of 100 lines of oldString. Atomic batches, auto-rollback. Learn more
📝 Your Neovim, embedded Real Neovim in a PTY — your config, your plugins, your LSP servers. Agent edits route through the same editor you use. Over SSH, in tmux, wherever.	🎚️ Mix-and-match models Haiku for exploration. Sonnet for code. Flash for compaction. The task router wires a different model to each job — cheap work goes to cheap models. 21 providers + any OpenAI-compatible API.
📱 Reach your forge from anywhere Hearth turns a running SoulForge into a remote agent. Telegram or Discord. Tap-to-approve for destructive ops, auto-redaction of secrets. Your code never leaves your host. Learn more	↶ Undo any turn Every prompt is a checkpoint. Ctrl+B / Ctrl+F walks history. Branching from any point rewrites the conversation AND restores files on disk. Learn more
📑 Tab-aware file claims Up to 5 tabs per project with independent model, mode, session, and checkpoints. Tabs see each other's claimed files and active agents. Git hard-blocks during cross-tab dispatch, partial commits are impossible. Learn more	🪝 Drop-in Claude Code hooks 13 lifecycle events (PreToolUse, PostToolUse, compaction, subagents). Reads your existing .claude/settings.json — no rewrites. Learn more
🔌 MCP-compatible Any Model Context Protocol server works out of the box. stdio, HTTP, SSE. Auto-reconnect, namespaced tools. Learn more	🧩 Skills Install domain-specific skills with Ctrl+S. Bun development, Three.js fundamentals, product marketing, whatever. Approval-gated, scoped per session.

And a lot more

• Steering — type while the agent works, messages inject mid-stream. More
• Lock-in mode — hide narration, show only tool activity and final answer
• Inline images — pixel-perfect images and animated GIFs via Kitty graphics protocol
• 24 themes — Catppuccin, Dracula, Gruvbox, Nord, Tokyo Night, Rose Pine, and more. Hot-reload custom themes. More
• Floating terminals — Ghostty-powered PTYs next to the chat
• Plan mode — research, write a structured plan, you approve, then execute. More
• Memory — persistent SQLite memory across sessions, scoped per project or global
• Pre-commit enforcement — git commit auto-runs lint + typecheck; fails block the commit
• 100 slash commands — full reference

Install

brew tap proxysoul/tap && brew install soulforge

macOS and Linux. Neovim and a Nerd Font auto-install on first launch.

Windows: not natively supported. Use WSL for now. Tracked in #57.

Other install methods

# Bun (global)
bun install -g @proxysoul/soulforge

# Prebuilt binary
# download from https://github.com/ProxySoul/soulforge/releases/latest
tar xzf soulforge-*.tar.gz && cd soulforge-*/ && ./install.sh

# Source
git clone https://github.com/ProxySoul/soulforge.git && cd soulforge && bun install && bun run dev

Get a key

Pick any one.

soulforge --set-key llmgateway sk-...         # one key for every major model, up to 30% off frontier
soulforge --set-key anthropic sk-ant-...      # or any individual provider you already have
soulforge                                     # launch, Ctrl+L to pick a model

All providers · Custom providers

How it compares

	SoulForge	Claude Code	Codex CLI	OpenCode
Codebase awareness	Live SQLite graph — PageRank + git co-change, blast-radius tags, per-turn personalization	File reads + grep	File reads + grep	File reads + grep
Cost tactics	Surgical reads, parallel shared cache, free V2 compaction, model-per-task router	Auto-compaction	Server-side compaction	Auto-compaction
Code intelligence	LSP → ts-morph → tree-sitter → regex, 33 languages, Mason installer (576+ servers)	LSP via plugins	—	LSP auto-load
Editor	Embedded Neovim — your config	—	—	—
Remote control	Hearth: Telegram, Discord	—	—	—
Multi-agent	Parallel dispatch + shared cache + edit coordination	Subagents + Teams	Multi-agent v2	Multi-session subagents
Hooks	13 events, Claude Code drop-in compatible	Hooks (PreToolUse, etc.)	—	—
Providers	21 + any OpenAI-compatible	Anthropic only	OpenAI only	75+ via Models.dev
License	BSL 1.1 (converts to Apache 2.0 in 2030)	Proprietary	Apache 2.0	MIT

_{Verified April 2026. Report inaccuracies.}

Real numbers

All from SoulForge's own codebase, on Claude Sonnet 4.6:

Rename a class across 8 files	19 steps, $0.228 (text edits) → 3 steps, $0.036 (rename_symbol)
Change a function return type	~100 lines of oldString/newString → ~5 tokens with ast_edit
Compact a 34-message session	V1 LLM summary: ~8k output tokens, 5-15s → V2: 0 tokens, instant
Post-compaction conversation	4.5M prompt tokens → 7.5k tokens (context utilization 6% → 4%)

Claude Code's Explore subagent averages ~$0.70 per 5-minute research run with Haiku. SoulForge matches it when you route spark to Haiku via the task router — with the added benefit of full repo-map context.

Try it

brew tap proxysoul/tap && brew install soulforge
cd your-project
soulforge

Then:

> rename AgentBus to CoordinationBus across the project
> run tests and commit

Pair a Telegram/Discord bot once with /hearth pair, then keep chatting from your phone — the session auto-syncs both ways.

Full docs at soulforge.proxysoul.com.

License

Business Source License 1.1. Free for personal and internal use. Commercial use requires a commercial license. Converts to Apache 2.0 on March 15, 2030.

_{Open-sourced March 30, 2026. Built by proxySoul}