I build native systems around coding agents: workbenches, skills, deterministic verification, memory, routing, and forensics.
Each project works alone. Together, they form a closed execution loop.
Choose the problem you want to solve, then follow that project's setup guide.
| Your task | Start here |
|---|---|
| Run parallel coding agents in one native macOS workbench | fibo (fibo.page) |
| Find, reuse, and audit skills for coding agents | claude-skill-registry for discovery · spellbook for workflows · skill-checkup to lint installed skills |
| Stop unverified completions and guard agent changes | vibeguard · receipt |
| Keep context across long-running agent work | remem |
| Route requests across 60+ LLM providers in Rust | litellm-rs |
| Track Claude, Codex, Cursor, and Grok quotas & costs | quotabar · ccstats |
Each repository documents its own installation, requirements, support, and license. The stack diagram below describes how the projects relate; it is not a shared installer.
WORKBENCH / 01
Native macOS workbench for coding agents (GPUI + libghostty): persistent daemon, live attention states, never lose a session. 
|
EXTEND / 02
Searchable discovery catalog for Claude Code & agent skills, paired with spellbook workflows. 
|
ROUTE / 03
High-performance Rust LLM gateway with 60+ runtime-wired providers through one OpenAI-compatible contract. 
|
ORCHESTRATE / 04
Governed fleets of parallel coding agents, powered by a Rust control plane. 
|
TRUST / 05
Scoped review rules, native Bash/Git hooks, and guardrails against unverified agent changes. 
|
REMEMBER / 06
Local-first, SQLCipher-encrypted memory for long-running Claude Code and Codex work. 
|
// OPEN_MODULE_BAY — Rust crates, verification probes, and agent skills
| Project | Role |
|---|---|
jsonrepair-rs |
Repair 30+ classes of malformed JSON from LLM outputs (150k+ downloads on crates.io) |
agent-sessions |
Streaming Rust parsers with byte-offset provenance across 25+ local coding-agent formats |
markitdown-rs |
4.5 MB Rust rewrite of markitdown (401/406 byte-identical, median 408× faster CLI + MCP server) |
ingot |
Native Rust + custom Metal MoE inference engine for Apple Silicon |
rekey |
Local Credential Authority for AI agents backed by macOS Secure Enclave |
rnk |
Declarative Rust TUI framework with React-like hooks and 45+ components |
palim |
Reversible JSON diff, RFC 6902 Patch, and RFC 7396 Merge Patch in Rust |
ccstats |
Local-first token and cost analytics engine across AI coding agents |
| Project | Role |
|---|---|
receipt |
Deterministic Stop hook blocking hallucinated "tests pass" / "committed" claims without proof |
paperdiff |
Line-by-line arXiv paper vs. GitHub code hyperparameter & seed consistency checker |
stealthprint |
Fingerprint anonymous or proxied LLMs via tokenizer, wrapper, context, and vision probes |
keyspoor |
Offline secret scanning across Git hooks, CI, and read-only MCP |
argus |
Static supply-chain admission scanner across eight package ecosystems and lockfiles |
verdict |
Pin agent skill capability contracts to catch post-install drift and runtime overreach |
mcp-mux |
Multi-account OAuth multiplexer for any MCP server backed by macOS Keychain |
| Project | Role |
|---|---|
spellbook |
Cross-runtime workflow skills for Claude Code, Codex, and multi-agent pipelines |
skill-checkup |
Audit installed SKILL.md folders for drift, broken references, and description gaps |
quotabar |
Menu bar quota window, reset pace, and subscription value monitor for Claude, Codex, Cursor & Grok |
dsh-desk |
Bundled desktop distribution for DeepSeek Harness with daily compatibility checks |
awesome-grok-bot |
Curated, shelf-tiered catalog of 6,600+ live Grok Bot shares |
awesome-goal-prompts |
Source-backed /goal contracts for long-running autonomous coding sessions |
peerscope · you-are-here · agent-patterns · patch-tournament · shipwise |
Focused skills for competitive teardowns, overnight handoff maps, patch grading, and open-source launches |
BEIJING · UTC+8 · USUALLY BUILDING SOMETHING THAT KEEPS AGENTS HONEST