xt-forge
Health Gecti
- License — License: MIT
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Community trust — 19 GitHub stars
Code Basarisiz
- spawnSync — Synchronous process spawning in bin/npx-xtforge.mjs
- process.env — Environment variable access in bin/npx-xtforge.mjs
- rm -rf — Recursive force deletion command in install.sh
Permissions Gecti
- Permissions — No dangerous permissions requested
Bu listing icin henuz AI raporu yok.
Local agent workbench for call-graph context, offline reverse engineering, CVE intel, and sub-1KB MCP routing. Built in Rust, Go, and stdlib Python.
xt-forge
Supreme local agent workbench for call-graph context, AST correctness auditing, normalized binary diffing, keyless CVE intel, and sub-1KB MCP routing.
Engineered in Rust, Go, and zero-dependency Python 3.11+ under XT-OSS by XtrComSu.
XT-Forge vs. Raw REA
Mounting a 229-tool reverse-engineering MCP server like raw REA (rea-agents mcp, ~3.5 MB tool catalog) directly into an agent harness floods the prompt window, triggers tool-discovery loops, and times out on multi-file audits, binary regressions, and dropped-connection recovery.
xt-forge solves this at both layers:
- Sub-1KB
--metaGateway (1,384 B/346tokens in Go,98.5%schema reduction): Routes localxt-forgeengines plus stateful upstream MCPs (Borer, REA, Craft Anything) through 4 deterministic primitives (xt_discover,xt_schema,xt_call,xt_batch) with crash-safe idempotency journaling ($XT_HOME/call_journal.json). - Direct Structural & Binary Analysis (
code_callers,code_audit,code_trace,re_diff): Replaces multi-turngrep,read_file, andobjdumploops with single-call AST, CFG, and disassembly analysis.
| Benchmark Tier | Workload | Raw REA Only (229 tools) |
xt-forge + REA (--meta) |
Key Advantage |
|---|---|---|---|---|
| Tier 1 · Find | Symbol definition + production callers vs test/doc decoys | 28.8s (8 tools, misses tuple / extra grep) |
PASS · 15.1s ev (code_callers) |
Separates production_callers from unresolved_candidates (test_or_shadow_only) in one indexed query |
| Tier 2 · Fix | Off-by-one retry budget bug fix + test verification | 52.9s (9 tools, 13,864 tok) |
PASS · 34.2s (7 tools, 9,106 tok) |
1.55x faster and 34% fewer tokens via direct symbol localization |
| Tier 3 · Recover | Stateful MCP job submission across mid-flight connection drop | FAIL (0/1) · Duplicate commit (2 submits) |
PASS (1/1) · 22.7s (0 duplicates) |
Crash-safe $XT_HOME/call_journal.json runs verify_tool before retry and recovers committed receipt |
| Tier 4 · Defects | Multi-file static correctness audit (3 bugs + clean decoys) | 72.3s (10 file reads) |
PASS · 28.8s (5.2s ev, 1 xtf call) |
2.5x faster — code_audit extracts all 3 AST defects (duplicate_aggregation, boundary_comparison, stale_cache) + full function_source in one call |
| Tier 5 · Trace | Active runtime path across __init__.py re-exports & config dispatch |
32.7s (9 manual file reads) |
PASS · 20.2s (7.1s ev, 1 xtf call) |
1.62x faster — code_trace resolves runtime.json → active_v2_handle → active_v2.py in one hop |
| Tier 6A · Src Diff | Behavioral regression localization across v1 vs v2 C source |
31.6s (5 tools, 4,650 tok) |
PASS · 20.5s (3 tools, 3,109 tok) |
1.54x faster and 33% fewer tokens |
| Tier 6B · Bin Diff | Stripped/compiled ELF binary regression (calc_v1 vs calc_v2) |
FAIL (120.0s TIMEOUT) · 12 tools |
PASS · 69.4s (5.2s ev, re_diff) |
re_diff normalizes instructions, annotates .rodata strings, and decodes GCC reciprocal division (10% -> 5%) automatically |
Why Use xt-forge
| Everyday Problem | What xt-forge Gives You |
Command / Guide |
|---|---|---|
| Bloated MCP catalogs eat 10K–40K+ tokens upfront | --meta gateway exposes 4 meta-tools (1,384 B Go / 1,007 B Python) across local tools + external MCPs |
xt-forge --meta |
Massive RE/IDE MCPs (ghidra, borer, craft-anything) break context limits |
Connect once (mcp connect), SHA-256 trust-pin, and route stateful stdio calls lazily with 1-call load_binary optimization |
xt-forge mcp connect ghidra |
100+ MB stripped ELFs (libroblox.so), sectionless binaries (e_shnum=0), or Android Bionic APS2 relocations OOM Ghidra or mislead tools |
Streaming RIP-relative lea string-anchor resolver (2.7s on 120.6 MB), O(1) windowed disassembly, PT_DYNAMIC sectionless ELF parser (5.6ms), and RSS-budgeted process_vm_readv live probe |
xt-forge re deep-elf target.so --json |
| Multi-file bug hunts burn 10+ file reads and miss subtle defects | One-call AST correctness auditor flags duplicate joins, boundary < vs <=, stale caches, and floor-div truncation with full function source |
xt-forge audit . --json |
| Re-exports and config dispatch tables hide the real handler | Cross-module active-path tracer resolves __init__.py aliases and JSON/TOML/YAML/ENV dispatch keys |
xt-forge trace dispatch_request . |
Comparing two binaries requires manual objdump diffing & magic-number math |
Normalized per-function instruction diffing, inline .rodata string annotation, and GCC/Clang reciprocal division decoding |
xt-forge diff v1.bin v2.bin --json |
| Regex grep misses callers, callees, and definitions | Incremental code_graph.db call graph + Go LSP multiplexer pack structural slices and separate test shadows |
xt-forge callers MyFunc . --json |
| Threat lookups stall behind API keys or rate limits | Keyless NVD, FIRST EPSS, CISA KEV, InternetDB, GreyNoise, and CIRCL fan-out | xt-forge intel cve CVE-2024-3094 |
| Agents repeat failed fixes or duplicate mutating MCP calls on timeout | Crash-safe $XT_HOME/call_journal.json (idempotency_key + verify_tool), negative memory, and goal verify |
MCP & Steering Guide |
Install on Linux, macOS, and Windows
Option 1: One-Line Native Install (Builds Rust + Go + Links PATH + Syncs 24 Skills)
Linux & macOS (Bash / Zsh):
# Piped one-liner (clones into ~/.xt-forge-src if run outside a repo):
curl -fsSL https://raw.githubusercontent.com/XtrComSu/xt-forge/main/install.sh | bash
# Or from a local clone:
git clone https://github.com/XtrComSu/xt-forge.git && cd xt-forge && ./install.sh
Windows (PowerShell 5.1+ / PowerShell 7+):
# Piped one-liner:
irm https://raw.githubusercontent.com/XtrComSu/xt-forge/main/install.ps1 | iex
# Or from a local clone:
git clone https://github.com/XtrComSu/xt-forge.git; cd xt-forge; .\install.ps1
Option 2: Direct GitHub npx / uv / pip Install
# Node.js (18+) directly from GitHub (installs xt-forge, xtf, /xtf, and 24 XTF-* skills):
npx --yes github:XtrComSu/xt-forge init
# Or install the npm tarball from GitHub Releases:
npm install -g https://github.com/XtrComSu/xt-forge/releases/download/v0.4.5-Damascus/xt-forge-0.4.5.tgz
# Python (3.11+) wheel from GitHub Releases or git:
uv tool install https://github.com/XtrComSu/xt-forge/releases/download/v0.4.5-Damascus/xt_forge-0.4.5-py3-none-any.whl
# or: pip install git+https://github.com/XtrComSu/xt-forge.git
Option 3: Prebuilt Release Archives (v0.4.5-Damascus)
Prebuilt platform archives, Python wheels, and SHA256SUMS.txt are attached to GitHub Releases (v0.4.5-Damascus):
xt-forge-0.4.5-linux-x86_64.tar.gz(includes prebuiltxtforge-nativeRust +xtforge-hubGo binaries)xt-forge-0.4.5-darwin-arm64.tar.gz(includes prebuiltxtforge-hubmacOS arm64 binary)xt-forge-0.4.5-windows-x86_64.zip(includes prebuiltxtforge-hub.exeWindows x86_64 binary)xt-forge-0.4.5.tgz(npm package tarball)xt_forge-0.4.5-py3-none-any.whl&xt_forge-0.4.5.tar.gz(Python wheel & sdist)
xt-forge --meta # stdio MCP gateway (xt_discover, xt_schema, xt_call, xt_batch)
xt-forge doctor # inspect workspace health, guard mode, and native engines
xt-forge skills # list all 24 bundled XTF-* workflow skills
Agent Self-Install Prompts
Want an AI coding agent (Claude Code, Codex CLI, OpenCode, DSH, Cursor) to install and verify xt-forge itself? Paste this prompt directly to the agent (see prompts/install.md):
Install
xt-forge(https://github.com/XtrComSu/xt-forge) for this OS (curl -fsSL https://raw.githubusercontent.com/XtrComSu/xt-forge/main/install.sh | bashon Linux/macOS,irm https://raw.githubusercontent.com/XtrComSu/xt-forge/main/install.ps1 | iexon Windows PowerShell, ornpx --yes github:XtrComSu/xt-forge init), verifyxt-forge,xtf,/xtf, and all 24XTF-*skills are installed, registerxt-forge --metain the active MCP config if one exists, and runxt-forge version --json+xt-forge doctor --json+tools/listonxt-forge --metato prove the setup works.
Lean MCP Setup + Ghidra, Borer & Craft Anything (Full REA Replacement)
Wire xt-forge once using --meta. Models discover tools on demand (xt_discover -> xt_schema -> xt_call or xt_batch), keeping initial prompt schema size at 1,384 bytes (346 tokens) in Go (1,007 B / 252 tok in Python) — 94.4% smaller vs Borer alone and 99.96% smaller across all connected upstreams — and spilling oversized responses cleanly to $XT_HOME/cache/artifacts/.
It natively connects, trust-pins, and optimizes Ghidra MCP (xtforge.adapters.ghidra_mcp, 43 MCP tools compatible with LaurieWired/GhidraMCP, pyghidra-mcp, and rea-agents — featuring persistent SHA-256 function/variable/prototype/comment overlays, live Ghidra HTTP bridge, inspect_elf_deep for e_shnum=0 & Android Bionic APS2 ELFs, resolve_string_anchors + slice_disasm_window for 120+ MB stripped binaries without Ghidra OOM, probe_live_process with smaps RSS zero-page protection, and inspect_rbxm_container), Borer (borer, 34 MCP tools + 111 borer:// resources for certified decompilation, function emulation, symbolic solve, devirtualization, and unpacking), and Craft Anything (craft-anything-cli mcp, 300+ tools) behind persistent stateful stdio sessions, 1-call path auto-load_binary optimization, borer:// shorthand URI expansion, and unified SHA-256 rug-pull checks — replacing heavy Node.js rea-agents setups completely:
# 1. Connect, injection-scan, and SHA-256 pin Ghidra, Borer, or Craft Anything once
xt-forge mcp connect ghidra --json
xt-forge mcp connect borer --json
xt-forge mcp connect craft-anything --json
# 2. One-call Ghidra & Borer execution (auto-loads binary + applies persistent C overlays)
xt-forge tool run ghidra rename_function '{"path":"/usr/bin/ls","old_name":"0x5620","new_name":"ls_entrypoint"}'
xt-forge tool run ghidra set_function_prototype '{"path":"/usr/bin/ls","function_address":"0x5620","prototype":"int ls_entrypoint(int argc, char **argv)"}'
xt-forge tool run ghidra decompile_function '{"path":"/usr/bin/ls","name":"ls_entrypoint"}'
xt-forge tool run ghidra inspect_elf_deep '{"path":"/usr/bin/ls"}'
xt-forge tool run ghidra resolve_string_anchors '{"path":"/usr/bin/ls","queries":["POSIX"]}'
xt-forge tool run borer query '{"path":"/usr/bin/ls","uri":"capabilities"}'
# 3. Or proxy any server directly with auto-refreshing dynamic toolsets
xt-forge mcp proxy --server ghidra
xt-forge mcp proxy --server borer
xt-forge mcp proxy --server craft-anything
{
"mcpServers": {
"xt-forge": {
"command": "xt-forge",
"args": ["--meta"]
}
}
}





3-Tier Workflow Skills (XTF-*)
xt-forge init syncs 24 workflow skills (8 Simple, 8 Complex, 8 Advanced) into ~/.agents/skills, ~/.dsh/skills, ~/.config/opencode/skills, and ~/.claude/skills so models pick exact commands instead of guessing:
| Tier | Count | Skills | Scope |
|---|---|---|---|
| Simple | 8 | XTF-RepoMap, XTF-CallGraph, XTF-CodeAudit, XTF-DispatchTrace, XTF-Notes, XTF-SecretScan, XTF-LogCompress, XTF-CVELookup |
Workspace maps, call graphs, multi-file AST correctness audits, active config/re-export dispatch traces, flash notes, secret/injection scans, log compression, keyless CVE lookup |
| Complex | 8 | XTF-BinaryTriage, XTF-BinaryDiff, XTF-Disasm, XTF-LSPNavigate, XTF-IOCIntel, XTF-BinaryPatch, XTF-Unpack, XTF-DeepResearch |
Offline ELF/PE triage, normalized binary diffing + .rodata string annotation + reciprocal math decoding, disassembly/CFG/RTTI, Go LSP stdio navigation, IOC enrichment, guarded patching, ASAR/APK/HAR unpack, parallel research |
| Advanced | 8 | XTF-Decompile, XTF-SwarmRE, XTF-MCPGateway, XTF-SessionRecovery, XTF-BugFix, XTF-CrashTriage, XTF-GoalLoop, XTF-MemoryEval |
XOR/stack string decoding, headless decompilation, parallel binary MapReduce & MCTS hypothesis solving, MCP rug-pull defense, crash-safe call journaling & idempotency recovery, verified bug fixes, minidump postmortem, bounded verification loops |
Full reference and trigger patterns live in docs/SKILLS.md.
Harness Integrations
Drop-in configurations, slash commands (/xtf --meta), and tailored setup walkthroughs are included for each major agent harness:
| Harness | Folder | Integration Highlights |
|---|---|---|
| DSH | integrations/dsh/ | Native /xtf --meta command prompt, mcp-patch.json, profile-aware cordis setup |
| OpenCode | integrations/opencode/ | opencode.json + native /xtf, /taste, /goal, /loop TypeScript plugin |
| Codex CLI | integrations/codex/ | Clean config.toml entry + compact AGENTS.md workflow rules |
| Claude Code | integrations/claude-code/ | Project .mcp.json, CLAUDE.md, and /xtf slash command integration |
Common Commands
# One-Shot Code Intelligence, AST Auditing & Active Dispatch Tracing
xt-forge callers should_retry_attempt . --json
xt-forge audit . --json
xt-forge trace dispatch_request . --json
xt-forge context . --query "verify_sig" --symbols --token-budget 1200
# Normalized Binary Diffing, Decompilation, Stripped 100+ MB ELF Anchors & Keyless CVE Intel
xt-forge diff bin/calc_v1 bin/calc_v2 --json
xt-forge re decompile payload.elf --symbol apply_tier_discount --json
xt-forge re deep-elf payload.so --json
xt-forge re anchors payload.so --target "bytecode version mismatch" --json
xt-forge re slice payload.so --start 0x36085e4 --json
xt-forge swarm map payload.elf --workers 8 --json
xt-forge intel cve CVE-2024-3094 --json
xt-forge re checksec payload.elf --json
# Connect Ghidra + Borer MCPs, Crash-Safe Recovery & Security Scanning
xt-forge mcp connect ghidra --json
xt-forge mcp connect borer --json
xt-forge session recover --json
xt-forge scan secrets .
Native Performance
Hot CLI paths run directly in Rust (xtforge-native) and Go (xtforge-hub), delivering 10x to 25x speedups over XT_FORCE_PYTHON=1 on cold subprocess invocations (median of 5 runs):
| Workload | Native Engine | Native (ms) | Python (ms) | Speedup |
|---|---|---|---|---|
scan secrets xtforge |
Rust | 28.2 | 327.2 | 11.6x |
route context xtforge --symbols |
Rust | 15.3 | 162.8 | 10.7x |
code index xtforge |
Rust | 6.9 | 122.9 | 17.9x |
codec entropy <elf> |
Rust | 7.7 | 197.5 | 25.8x |
re checksec <elf> |
Go | 8.8 | 120.6 | 13.7x |
re capa <elf> |
Go | 18.3 | 313.9 | 17.2x |
intel sources |
Go | 6.0 | 109.3 | 18.2x |

Benchmarks & Reproducibility
Raw multi-run benchmark logs and evaluation outputs are committed under benchmarks/:
| Suite | Scope | Measured Result | Command / Raw Data |
|---|---|---|---|
| Context Packing | context_tasks.json (20 tasks) + held-out context_tasks_unseen.json (8 tasks) |
20/20 (608.4 mean tok, 88.8% cut vs 3/20 map) · 8/8 unseen (566.4 mean tok) |
python3 scripts/bench_context_pack.py --json |
| Discovery Cost | Direct vs --meta on 4, 50, and 229 tools |
Direct wins 1-turn on 4 tools (450 vs 1,591 tok); --meta saves 93.1% (50) & 98.5% (229, 4.2 ms first call) |
python3 scripts/bench_discovery_cost.py --json |
| External CPython Stdlib Audit & Call Graph | 66 unmodified CPython 3.12 files (json, urllib, email, http, xml, logging, sqlite3, 3 reps) + 8-category defect matrix |
0 false positives across 198 audited stdlib file-runs (0.0% FP rate) · 8/8 defect categories (1.0 recall, 1.0 precision) · 12/12 call-graph hits |
python3 scripts/bench_external_repos.py --reps 3 · raw JSON |
| 6-Tier Live Harness Suite vs REA | Live A/B runs (Find, Fix, Recover, Defects, Trace, Regression Src/Bin) with rea-agents (229 tools) |
34 recorded live runs committed (--self-test verifies fixtures/graders offline; omit --self-test to run live harness comparison) |
python3 scripts/bench_three_tier.py --harnesses dsh --arms A,B --reps 3 --user-harness --dsh-keep-mcps mcp-rea · raw JSON · CSV |
Documentation
- Field Guide & Architecture Walkthrough
- CVE Research and Reverse Engineering Deep Dive
- 24 Bundled 3-Tier Skills Reference (
XTF-*) - MCP Meta-Gateway and Steering
- Rust + Go Hybrid Runtime Details
- RE Workbench Details
- Changelog (
v0.4.3-Damascus) · Architecture Release Notes
Tests
cd native/xtforge-rs && cargo test && cargo build --release && cd ../..
cd native/xtforge-go && go test -race ./... && go build -o xtforge-hub ./cmd/xtforge-hub && cd ../..
python3 -m unittest discover -s tests
node --test plugins/xt-native/test/*.test.mjs
MIT Licensed · XT-OSS · xtrcom.xyz
Yorumlar (0)
Yorum birakmak icin giris yap.
Yorum birakSonuc bulunamadi