prompt-caching

See what Claude Code and Codex actually send to the API — and what each part costs.

A curated list on AI token economics: what tokens cost, where they get wasted, and how to cut the bill. Tools, benchmarks, papers, and copy-paste configs for the token economy of LLMs and coding agents.

Local-first MCP tools and dashboard for investigating Codex token usage, credits, costs, caching, and thread patterns.

The Multi-Agent Reasoning framework creates an interactive chatbot where AI agents collaborate via structured reasoning and Swarm Integration for optimal answers. Simulating a team that discusses, debates, and refines responses, it enables complex problem-solving and precise results. Now with Prompt Caching to reduce latency and costs.

AI coding agent with one Python core and three front-ends — headless CLI, Textual TUI, and an Electron desktop. Works with any OpenAI-compatible API, with risk-tiered permissions, event-sourced replayable sessions, and a fail-closed OS-level sandbox.

Claude Opus 5 API (claude-opus-5): model id, per-million-token pricing, cache write and cache read rates, streaming and Anthropic-style requests through an OpenAI-compatible gateway.

Drop-in prompt-caching fixes for the LLM agent harness you use. Point your AI coding agent at this repo and it ships the patches.

Local-first AI companion daemon — an agent harness with its own long-term memory. Remembers you, grows with you, extends and repairs itself. 跑在你自己电脑上的 AI 搭档守护进程：会记住你、随你成长、能自加技能、能自我修复。

A Complete faster runtime for coding agents. Optimize the journey, not the hop. Make coding agents 25% faster and 30% cheaper on average while keeping the quality same or more. Same Task, Same Quality, Faster and Cheaper.

The open engine behind Aedelgard: a persistent, self-hosted AI agent with a verbatim memory palace and a model-agnostic brain. Swap Claude, Gemini, OpenAI, Mistral, DeepSeek or a local model without losing the mind.

Hermes Agent context engine: the main model writes the compaction handoff on its cached prompt prefix

A minimal coding-agent harness built to measure what context management actually costs: constraint survival after compaction (8%→83%), default-FAIL verification against false completions, and KV-cache prefix design (21%→90%). Stdlib-only, real deepseek-chat runs

Tool schemas get re-prefilled on every LLM request. ContextCache compiles them into a KV cache once, stores it on disk and reuses it across users and sessions — TTFT stays near 200ms with 50 tools instead of 5.6s (Qwen3-8B, RTX 3090 Ti).

Compression with a quality contract — cache-aware, causally-pruned LLM context compression for agentic runtimes, certified non-inferior across 7 domains. Works with any SDK.

TypeSafe (Jev) skill routing for Hermes Agent: names the one skill worth loading, before the model call. Opt-in, stdlib only, ~$0.001 per routed turn.

MCP C++ project indexer for fast symbol, module, and source-range navigation in large C++20 codebases.

Agentic-AI framework w/o the headaches

Agent Skill that caches LLM image descriptions as XMP metadata inside image files, reducing token usage by ~92% on repeated reads. Works with 30+ compatible agents.

Universal cli2api gateway for Tencent CodeBuddy Code — OpenAI/Anthropic/Gemini/Codex formats, multi-key cache-aware rotation, vision, + Hermes Agent plugin. Python stdlib only, zero deps.

Architectural Moat: JIT-JEV Context OS — Epistemic runtime & JEV System 1 context gate for AI agents (L0 SQLite WAL <3ms, Epistemic Invariants I1–I10, CERN Zenodo DOI: 10.5281/zenodo.22649542)