Resources
A curated list. Every link was checked on 2026-11-03. Inclusion is not an endorsement: each project’s savings figures are its authors’ own claims, and we have not reproduced them. Benchmark on your own tasks and track a quality metric next to cost (see Agent Tokenomics).
Start here
Section titled “Start here”| Resource | Why |
|---|---|
| microsoft/TokenEconomizer | Experience-driven cost control for coding agents: micro-skills, tool pruning, loop guard. Paper: arXiv:2610.22732, project page. |
| Optimize AI credit usage in VS Code | Official practical guide. |
| Diagnose prompt caching with the Cache Explorer | Official: find the first prefix divergence between requests. |
| Azure OpenAI prompt caching | Minimum length, cached_tokens, prompt_cache_key, retention, billing and data-residency rules for Azure OpenAI. |
| OpenAI prompt caching | Cacheable prefix rules, settings that invalidate the cache, diagnostics. |
| Usage-based billing for individuals | How AI credits are charged. |
| github/awesome-copilot | Community instructions, agents, and skills for Copilot. |
| pleasedodisturb/awesome-llm-token-optimization | Another curated list of strategies, tools, and papers. |
Reduce tool output and command noise
Section titled “Reduce tool output and command noise”| Project | What it does |
|---|---|
| rtk-ai/rtk | CLI proxy that condenses output of common dev commands before the model sees it. |
| headroomlabs-ai/headroom | Compresses tool outputs, logs, files, and RAG chunks before they reach the LLM. |
| gglucass/headroom-desktop | Desktop app wrapper around the same idea for coding agents. |
| Dave-London/Pare | Structured, token-efficient MCP servers for git, test runners, npm, Docker. |
| claw-compactor | Multi-stage, reversible compression pipeline with code-aware analysis. |
| optscale-kompressor | Deterministic prompt compression for coding agents in one Rust binary. |
| fkiene/llmtrim | Local proxy that trims prompts, history, and tool output in API requests. |
| Paritok-official/paritok-4b-v1 | Non-destructive compression gateway for coding agents. |
Compress prompts and data
Section titled “Compress prompts and data”| Project | What it does |
|---|---|
| microsoft/LLMLingua | Prompt compression research (LLMLingua, LLMLingua-2). Paper: arXiv:2310.05736. |
| jia-gao/leanctx | Drop-in prompt compression SDK for production apps (LLMLingua-2 based). |
| 3DAgentWorld/Toolkit-for-Prompt-Compression | Toolkit collecting prompt-compression methods. |
| ZongqianLi/Prompt-Compression-Survey | Survey of prompt-compression methods (NAACL 2025). |
| sriinnu/clipforge-PAKT | Lossless-first compression for JSON, YAML, CSV, Markdown. |
| ISON-format/ison | Token-efficient JSON alternative for LLM data. |
Terse style and measured evaluation
Section titled “Terse style and measured evaluation”| Project | What it does |
|---|---|
| JuliusBrussee/caveman | Skill/proxy that makes agents answer tersely. |
| danielle34/cavewoman | Evaluation framework measuring what input vs output compression actually saves. |
| YangJunMan/claude-code-token-savings-benchmark | Reproducible benchmark comparing several token-saving tools. |
Retrieve less code and keep memory small
Section titled “Retrieve less code and keep memory small”| Project | What it does |
|---|---|
| dotnetpower/code-context (Marketplace · Track A guide) | VS Code native Language Model Tools + shared local Rust trigram index (cctx) with <8 KiB inline-budgeted reads, BPE-formula instructions, 15.20× faster warm retrieval, and live status-bar/cctx usage AIC/USD savings telemetry. |
| Graphify-Labs/graphify | Turns a folder (code, docs, PDFs, images) into a queryable knowledge graph. |
| jgravelle/jcodemunch-mcp | Symbol-level retrieval: fetch the function you need, not the whole file. |
| redhat-et/ripwire | CLI + MCP server to find relevant code without reading the repo. |
| yamadashy/repomix | Packs a repository into one AI-friendly file (use selectively; big packs cost tokens). |
| radimsem/remindb | Portable agent memory in a SQLite file. |
| cytostack/openwolf | Project memory across agents, with token accounting from transcripts. |
| fajarhide/contexere | One place for skills, keys, and memory across agents. |
Audit tokens, tools, and cost
Section titled “Audit tokens, tools, and cost”| Project | What it does |
|---|---|
| tokprof (this repository) | Breaks each call’s input tokens down by system prompt, history, user prompt, and tool content without code changes; early development. |
| alexgreensh/token-optimizer | Finds “ghost tokens” and context-quality decay in agent sessions. |
| activeing123/mcptoon | MCP gateway CLI aimed at tool discovery and context compression. |
| ccusage/ccusage | Usage and cost reports from local Claude Code logs. |
| Maciek-roboblog/Claude-Code-Usage-Monitor | Real-time usage monitor with predictions. |
| junhoyeo/tokscale | Token usage across several AI coding agents from the terminal. |
Observability and gateways (for agents you operate)
Section titled “Observability and gateways (for agents you operate)”| Project | What it does |
|---|---|
| langfuse/langfuse | Open-source tracing, evals, and cost tracking for LLM apps. |
| Arize-ai/phoenix | AI observability and evaluation on OpenTelemetry (docs). |
| Helicone/helicone | LLM observability platform and OpenAI-compatible gateway (docs). |
| LangSmith | LangChain’s hosted tracing and evaluation platform with token and cost breakdowns (SDKs: Python, JS). |
| Portkey-AI/gateway | AI gateway with guardrails; Portkey’s platform adds monitoring (docs). |
| openlit/openlit | OpenTelemetry-based tracing and evaluation for agents. |
| AgentOps-AI/agentops | Agent monitoring and LLM cost tracking SDK. |
| BerriAI/litellm | Gateway with cost tracking and routing across 100+ providers. |
| maximhq/bifrost | LLM / MCP / agent gateway with failover. |
These map to stage 1 (Diagnose), which compares Phoenix, Langfuse, LangSmith, Helicone, and Portkey. They give you the per-request log that the example script consumes.
Context engineering and spec-driven workflows
Section titled “Context engineering and spec-driven workflows”| Project | What it does |
|---|---|
| github/spec-kit | Toolkit for spec-driven development: plan first, then implement. |
| Fission-AI/OpenSpec | Spec-driven development for AI coding assistants. |
| shanraisshan/claude-code-best-practice | Practices for agentic engineering with Claude Code. |
| microsoft/vscode-copilot-chat | Source of the VS Code Copilot Chat extension. |
Papers
Section titled “Papers”The full ranked list (by venue tier, then citations) is on the Papers page. Highlights:
| Paper | Takeaway |
|---|---|
| Evaluating AGENTS.md (arXiv:2602.11988) | Context files did not generally improve task success and raised cost by over 20% on average; useful for non-standard practices. |
| Lost in the Middle (arXiv:2307.03172) | Models use information in the middle of long contexts less reliably than at the start or end. |
| TokenEconomizer (arXiv:2610.22732) | 18.4–30.9% lower mean cost on SWE-Bench across 5 agent–model combinations, as reported. |
| LLMLingua (arXiv:2310.05736) | Prompt compression with small language models. |
| tiktoken | Tokenizer used by our token counter. |
Suggest a resource
Section titled “Suggest a resource”Open an issue or PR with the link, a one-line description in your own words, and how you verified it exists.