Paritok compresses the tools, files, and history your coding agent sends. Save up to 85% on your token bill and run 3× longer sessions. Two commands, nothing lost, fully local.
Loading comments…
Project Info
Product Keywords
Paritok is a local, self-hosted compression layer that sits between your coding agent and its language model upstream. It intercepts the tool schemas, file reads, and conversation history your agent sends on every turn, then rewrites them into a compact form before they reach the API. The response comes back untouched, and any compressed content can be expanded on demand. The result: up to 85% lower token bills and sessions that run roughly 3× longer, all with a single environment variable as the entire integration.
The entire integration is a single export: ANTHROPIC_BASE_URL=http://127.0.0.1:8080. Point your agent at Paritok, and it transparently handles compression from there. No SDK changes, no plugin installs, no configuration files to maintain.
Paritok attacks token waste from three angles simultaneously. First, it trims tool schemas from roughly 29K tokens to 8K per turn by keeping only relevant tools and stubbing the rest. Second, it compresses file reads and tool output to about 26% of their original size using a 4B model that preserves identifiers, paths, and errors. Third, it summarizes stale conversation history once your context budget fills, leaving recent turns untouched.
Nothing is ever truly discarded. When the agent needs the exact bytes of a compressed file or tool result, it can call read_original(ref) or gateway_search_tools(q) to fetch them locally — without burning an extra API turn. This makes the compression lossy on the wire but fully recoverable when it matters.
Because every file read stays in history and gets re-sent each turn, compression pays off repeatedly. A 1-turn session saves about 25%, but by turn 20 the savings reach 63%, and the ceiling depends on your setup — default configurations with ~40 tools can hit 72% savings per turn.
Lossy on the wire, recoverable when it counts.
Most compression tools force a permanent trade-off between token savings and data fidelity. Paritok sidesteps that entirely by making every compressed piece retrievable on demand, locally, without consuming an API call. This means you can aggressively shrink what goes upstream while keeping the agent's ability to access exact source bytes intact. The compounding effect is the real differentiator — the longer your session runs, the more you save, which directly enables the 3× longer sessions that heavy agent users need.
You're running long coding-agent sessions and watching token costs climb, or you're hitting context-window limits mid-task. Paritok is especially compelling if you already self-host your tooling and want a drop-in solution that requires no agent modifications. It's also worth a look if you're using any OpenAI-compatible upstream and want to cut spend without sacrificing the ability to retrieve exact file contents when needed.
Other tools you might consider
The moment an agent needs to deploy something, it slams face-first into a wall built for humans. Today we're rolling out Temporary Accounts on Cloudflare Workers. Any agent can now run wrangler deploy — temporary and get a live Worker in seconds.
GitHits gives coding agents access to the open-source code your app depends on. Get real implementation examples, dependency source navigation, package inspection and documentation. Agents can grep and read your codebase. They can't grep and read the open-source code your app depends on. That's where they start guessing, retrying, and looping. GitHits builds a version-aware index on demand. Agents can search, navigate, and inspect the code behind their dependencies. CLI: npx githits@latest init
Same AI. 5x the tokens. Coworker provides deep company context and automatically routes to the right model for every task. More chat, cowork and code with the same spend.
Octopoda is an open source infrastructure layer that gives AI agents persistent memory, automatic loop detection, and full observability. Agents forget everything between sessions. Octopoda fixes that with a remember/recall API that survives restarts, crashes, and deployments. The loop detection system monitors 5 signals to catch agents stuck repeating themselves before they burn your API budget, with real-time cost estimation showing exactly how much each loop wastes. A built-in dashboard shows every agent's health score, memory explorer with version history, audit trail logging every decision with reasoning, and a timeline replay that lets you scrub through everything your agent did step by step. Works with LangChain, CrewAI, AutoGen, and OpenAI Agents SDK with one-line integrations. Runs locally with SQLite or connects to cloud with one environment variable. Free tier, MIT licensed. 177 signups, 120 GitHub stars, zero marketing spend.
Maker
blueprint_b
Alternatives
Loading comments…
