I Built an AI Coding Agent That Already Knows Your Codebase Before It Reads a Single File

Most AI coding agent starts the same way.
You give it a task, It greps, It reads a file, It greps again, reads three more files then Half its token budget gone, and it still doesn't understand the full architecture.
I spent months watching agents burn through context and time just to figure out where things are. Not because the models are bad. Because they're flying blind. They see files. They don't see systems.

So I built one that does.
What SoulForge Actually Does Differently
When you open a project, SoulForge parses every file across 30+ languages, extracts every symbol and import edge, and builds a live dependency graph backed by SQLite.
Before the agent touches anything, it sees a Soul Map, every file ranked by importance, every exported symbol indexed, every dependency arrow drawn:

The graph updates in real-time. Edit a file, the map recalculates. The agent always sees the current state.
This isn't context stuffing. It's understanding.
Tools That Query the Graph, Not the Filesystem
Most agents get grep and read_file and hope for the best. SoulForge gives the agent tools that answer from the graph directly, zero file reads, zero LLM tokens:

One graph query replaces a dozen grep-read cycles.

The agent does less work, makes fewer mistakes, and you trust the result more.
Your Conventions, Loaded Once
Drop a SOULFORGE.md in your project root. Architecture notes, naming conventions, toolchain preferences, and it's injected into every prompt.
Already using Claude Code, Cursor, or Copilot? SoulForge loads instruction files from 10 formats: CLAUDE.md, .cursorrules, AGENTS.md, .github/copilot-instructions.md, and more. Your conventions follow you across tools.
Plus installable skills for domain-specific knowledge. The agent knows how your codebase works, not just what's in it.
4-Tier Code Intelligence
For precise operations, SoulForge routes through four tiers: LSP → ts-morph → tree-sitter → regex.
The agent gets the full LSP surface, definitions, references, workspace rename, call hierarchy, diagnostics. When LSP isn't available, it falls through gracefully. Every language always gets the best available intelligence.
This powers compound tools that do the complete job in one call:
One tool call. Not twelve.
Parallel Agents That Actually Coordinate

Big tasks fan out to parallel agents through a shared bus:
Up to 8 agents, 3 running concurrently. One agent reads a file, every other agent gets it instantly. Edit ownership prevents race conditions. A dependency DAG ensures code agents start with research already done.
The agent decides when to dispatch. Small tasks stay direct. Complex tasks fan out automatically.
Your Actual Neovim, Embedded
This was non-negotiable for me. I wanted to co-edit alongside the agent, not just watch a chat window.
SoulForge embeds real Neovim via msgpack-RPC. Your config loads. Your plugins work. LazyVim, Mason, treesitter, Catppuccin, all of it. The AI and your editor share the same LSP connection. When the agent renames a symbol, you see it in the buffer. When you edit, the agent sees it.
It works over SSH. tmux, screen, remote servers. No Electron, no browser. Terminal-native.
10 Providers, Per-Task Routing
LLM Gateway, Anthropic, OpenAI, Google, xAI, Ollama (local), OpenRouter, Vercel AI Gateway, a local proxy, and any custom OpenAI-compatible API.
Per-task model routing sends different tasks to different models automatically. Opus for coding accuracy. Haiku for cheap search queries. Sonnet for planning. Cost visible per task, per agent, per model.
You see exactly what you're spending.
Context That Doesn't Rot

This is where every AI tool breaks down. Hour-long sessions. Context fills up. Every API call sends 100K tokens of stale content.
SoulForge tracks working state deterministically from tool calls. When context fills up: serialize, compress, continue. 80% token reduction. Zero LLM cost. Sessions save incrementally, crash, close, resume from exactly where you left off.
Safety That Isn't Optional
Approval gates on every destructive action, rm -rf, git push --force, edits to .env or credentials. Individually prompted. No "Always Allow" button.
Forbidden files blocked across every tool. The LLM never sees your secrets. Pre-commit gate auto-runs lint and typecheck before any git commit.
The Numbers
BSL 1.1 — free for personal and internal use.
Get Started
Homebrew:
BUN:
Then just run: soulforge
Docs: soulforge.proxysoul.com
Source: github.com/ProxySoul/soulforge
If you've ever wished your AI coding agent understood what it was touching before it touched it, try it.
Disclaimer: SoulForge is in early release, if you face any issues don't hesitate to raise an issue or contibute <3
ProxySoul
#ai #claudeCode #claude #CLI #Agent #Neovim #OpenCode

