Every step a coding agent takes, it re-sends its entire working history to the model: every install log, every compiler warning, every progress bar. On a long task the model re-reads the same build output dozens of times and you pay for every read. The launches below are all aimed at that one fact, from several different angles.
Compaction at a clean break
Catnap is a Claude Code mod that lets Claude compact its own conversation and carry on. Claude Code normally compacts when the context window fills up, usually mid-task; catnap gives Claude a heads-up at 50% instead, so it finishes the piece it is on, writes itself a handoff note, compacts, and carries on without you typing anything. Moving the interruption from "whenever it happens" to "at a clean break" costs nothing and saves the thread.
One program instead of twenty tool calls
Code Mode MCP attacks the other half of context bloat. It is a Claude Code plugin where Claude writes one JavaScript program that calls your MCP tools in a sandbox, filters and joins the data there, and sends back only the answer. Instead of twenty result sets landing in the conversation, one answer does. The author measured the same correct answers at lower cost per task.
Compression that keeps the cache warm
Kompressor is deterministic prompt compression for coding agents in a single Rust binary. The key property is determinism: the same input compresses to the same bytes every turn, so the prompt cache keeps working. Every removed detail stays one retrieval away rather than being lost. Non-deterministic compression saves tokens and silently destroys your cache hit rate, which is why this one insists on being boring.
Knowing what any of it costs
Token Barometer tracks list prices for LLM APIs across thousands of models and hosted offerings, normalised to USD per 1M tokens, with full price history and a cloud-versus-local cost calculator. Each row shows where the price was read from and a confidence value, and data is refreshed daily from vendor pricing pages, APIs and aggregators. You cannot optimise a bill you cannot read.
Built for the model you can actually host
Reika is a coding agent CLI designed around small local models first, so it is careful with context, fails more gracefully and says when it is stuck. It does not make a small model smarter; it makes working with one less frustrating. Frugality by design beats frugality bolted on after the invoice arrives.
A language whose compiler answers back
Tov is a programming language for coding agents: TypeScript syntax that compiles to native binaries, with no any, no null and no double-equals, separate integer and float types, and calls that can fail marked try. After every edit you run the checker and apply each fix marked safe. When the compiler hands back the edit that fixes the problem, the agent spends fewer turns guessing.
Running generated logic safely, in-process
Stof runs AI-generated logic inside your own app: it is JSON with functions, so one document holds both the data and the logic that works on it, executed in a sandbox in-process in Rust, JavaScript or Python. Any JSON document is already valid Stof, and you can export back to JSON, YAML or TOML at any time. Keeping generated code in-process removes a whole class of round trip.
Finding out where your RAG actually breaks
RAG Evaluation Starter is a minimal Python script that evaluates a RAG pipeline against a golden set, reporting five metrics in one run starting with retrieval Recall@K — did the right document chunk come back in the top results? There is no SaaS, no signup and no mandatory API key. Retrieval quality is a context problem in disguise: pull the wrong chunk and the window fills with noise you paid for.
Profiling instead of tracing
Agency is an open-source platform from the UC Berkeley Sky Computing Lab for running, debugging, profiling and controlling AI agents. Model traces only tell part of the story, so it asks what an agent is waiting for and where time is being spent. Wall-clock profiling finds the waste that token counts hide, because a cheap step that blocks for ninety seconds is not cheap.
Measuring whether it worked at all
ReliaMesh is open-source reliability infrastructure for agents. It ingests structured outcomes, detects reliability deterioration and tracks recovery without collecting prompts, responses or tool content — the event contract rejects them. An API returning HTTP 200 does not establish that an agent completed its task. Cutting tokens is only worth doing if the task still completes, which is why measurement belongs in the same conversation as compression.
The trade to keep in mind
Every one of these trades some fidelity for some cost, except the measurement tools, which trade effort for certainty. The useful order is: measure first, then shrink the obvious waste, then compress what is left — and keep an eye on whether the task still finishes.