Small, tested, Apache-2.0 tools from running a fleet of Claude Code agents and local LLMs on our own hardware every day. Mostly standard-library Python. Each repo stands alone, has CI, and has a v0.1.0 release.
Measure Claude Code usage and spend fewer tokens for the same work.
| repo | what it does |
|---|---|
| pacekeeper | Claude Code usage pacing: 5h and 7d meters from the status line, %/hour targets, budgets, spend by project, and a wind-down hook. |
| turnthrift | Fewer turns, shorter replies: burst runs many reads in one call; hooks nudge batching and brevity in Claude Code. |
| leanagents | Route Claude Code subagents to small-prompt agent types and cheaper models automatically. |
| autocontext | Let an agent pick its own context window and autocompact threshold from measured cost and re-reads. |
| hushwire | Cut cross-session agent messages: status-only gate, digests, message stats. |
| escalator | Cheap-first patch pipeline: a local model drafts, a gate checks, a reviewer reviews, a stronger model fixes on red. |
| qsum | Extractive digests of long files through a local OpenAI-compatible endpoint. |
| codeindex | Fast local code index for agents to find symbols and files without an LLM. |
Run, schedule and hand over many long-running Claude Code sessions.
| repo | what it does |
|---|---|
| supra-orchestrator | Run a fleet of Claude Code agent sessions from one orchestrator: order ledger, cadence, escalation. |
| fleetwarden | Guardrail hooks and a session launcher for running many Claude Code sessions: cadence, agent routing, protected dirs, rm and wait guards, intent injection. |
| tickwise | Adaptive tick intervals for long-running agent sessions: slow down when nothing changes, speed up on change. |
| diplomat | One idle agent session per machine: watch, triage alarms, relieve load, CPU efficiency. |
| opsledger | Turn a big multi-part order into deduped, deconflicted asks with evidence-backed status and a percent-complete ledger. |
| oldest-asks | Find what a person asked their agents for that was never delivered, oldest first, with at most 5 Haiku workers. |
| landbot | Local merge queue for AI-agent branches: batch, gate, bisect, land. No CI service needed. |
| billet | Billet turnover for long-running Claude Code sessions: performance-based degradation check, verified turnover brief, clean successor. |
Keep agents on your stated intent, verify that work is done, and get stuck agents moving.
| repo | what it does |
|---|---|
| intentpin | Pin the person's own words into every Claude Code session start and compaction, newest wins. |
| anti-lobotomy | Audit a Claude Code setup for silent context loss: compactions, truncated reads, silent model rewrites, rule bloat and alarm noise. |
| donegate | Claude Code Stop hook: no 'done' without a test run after the last edit; replies kept short. |
| zoomout | Step back when an agent is stuck: how top practitioners would approach it, then continue, pivot or simplify. |
| freshlook | Escalate a project to a fresh, independent auditor that sees only code and intent; ranks complexity to remove. |
| breakthroughs | 3-10 cheap agents brainstorm anything that could work on a project; one stronger agent scores reliability. |
Run local LLM servers on your own hardware without cooking it.
| repo | what it does |
|---|---|
| llm-rig | Run local LLM servers in libvirt VMs with GPU passthrough: bring up, provision, status, benchmark. |
| vmlease | Lease-based libvirt VM orchestration for agent fleets: agents lease, a reaper reclaims, no hand virsh. |
| gpu-handoff | Hand a passthrough GPU between a libvirt VM and the Linux host at runtime, no reboot: precheck, dry-run, rollback. |
| modelstage | Staged rollout of local LLM model files (GGUF) across machines: sha-verified, health-gated switch, automatic rollback. |
| gsched | GPU airtime scheduler with leases, plus a queue wrapper that books GPU time. |
| jobq | One door for heavy jobs across nodes, each run in its own cgroup slice with limits and a queue. |
| rigwatch | Homelab GPU and VM safety: temperature guards, VM and GPU leases, RAM guard, heavy-job enforcer, and a boot check for LLM containers that started without CUDA. |
| nascare | NAS care on Linux: heat policy, spin-down, SMART and pool health, ZFS error-growth alerts. |
| corsair-fans | Temperature-driven fan control for a Corsair Commander Core on Linux, using the liquidctl CLI. |
Reliable PDF, Word and Excel operations for LLM tool layers, plus grading and test data.
| repo | what it does |
|---|---|
| pdfops | Python library of PDF operations for LLM agents: extract text and tables, merge, split and fill forms, with JSON schemas for tool calls. |
| docxops | Word .docx operations for LLM agents in Python, including a real table of contents, with JSON-schema tool definitions. |
| xlsxops | Excel .xlsx operations for LLM agents in Python (openpyxl), with JSON-schema tool definitions and safe saves. |
| officecheck | Deterministic grader for .docx, .xlsx, .pptx and .pdf outputs: a JSON spec goes in, pass or fail per check comes out, no model judge. |
| fakepii-docs | Synthetic test corpus generator for PII masking: fake personal data in docx, xlsx, pptx, pdf and csv, with ground truth and a leak scorer. |
Contributions: see each repo's CONTRIBUTING.md. License: Apache-2.0.