justokenmax

jusTokenMax

Keep your coding agent under a token (and cost) budget. jusTokenMax shrinks every expensive thing before it reaches the model's context — attachments, logs, JSON, notebooks, CSVs, diffs, and the files you read — so the same work costs a fraction of the tokens.

Built by Kashi (linkedin) and Rajan (linkedin), founders of Kalmantic — jusCode.co. MIT licensed. Contributors: Arbaz (linkedin), CTO of LineupX.

---

TokenMax under a budget

Coding-agent bills are driven by input tokens — the PDFs, logs, API responses, diffs, and source files that pile into the context window. jusTokenMax caps that: it intercepts each heavy input and replaces it with a faithful, far cheaper equivalent, before it costs you a token.

• Compress everything that bloats context — PDFs → Markdown, images downscaled, logs/JSON/notebooks/CSVs/diffs digested, whole-file reads replaced by symbol lookups. Typical reductions 56%–99% (measured, below).
• Stay under a budget — point it at the things you feed your agent and the per-task token cost drops by roughly the same amounts; pair with terse output and chat-branch to also cap what the agent writes and re-reads.
• Reversible & safe — every original is cached by content hash (justokenmax retrieve brings it back); secrets and base64 blobs are masked on the way through.
• Works where you work — automatic in Claude Code (a Read hook swaps heavy files for cheap artifacts in place), and available to any MCP agent (Codex CLI, OpenCode, Cursor, …) plus a plain justokenmax CLI.
• Zero dependencies, fully auditable — deterministic heuristics, no trained model, no network. Every transform is readable Python.

What you feed it	Typical reduction
PDF spec / paper	−56% (real PDFs)
Verbose build/CI log	−99%
Large JSON / API response	−99%
Jupyter notebook	−99%
CSV (thousands of rows)	−99%
Git diff (lockfile churn)	lockfile → 1 line
Finding a symbol vs reading the file	−97%

---

Quickstart — try it

Fastest path, inside Claude Code (run these one at a time — one slash command per prompt):

1. /plugin marketplace add https://github.com/Kalmantic/jusTokenMax.git
2. /plugin install justokenmax@justokenmax
3. /reload-plugins

Now reading PDFs / logs / JSON / CSV / notebooks / diffs is compressed automatically. (The hook calls the justokenmax CLI — see Install to add it; or just have Node and it auto-provisions.) Prefer the CLI or another agent? justokenmax install registers it for Codex / OpenCode / Cursor too.

👉 Full 5-minute hands-on with a real dev task and on/off measurement: docs/try-it.md.

In a real development loop — one pass through a small website project's inputs (8 source modules, a 10k-line package-lock.json, a 5,000-row CSV, a noisy build log): 259,819 → 103,745 tokens (−60%), measured with a real tokenizer — and it compounds as the agent re-reads files while editing. Reproduce it step by step in docs/try-it.md.

---

Built for casual users and enterprise — from day one

The same tool serves a solo developer and a regulated enterprise, by design — not as an afterthought:

• Casual users get a one-command setup (justokenmax install) and then it just works, automatically: free, MIT, no account, no signup, sensible defaults. Your agent simply gets cheaper.
• Enterprise gets a tool that's safe to adopt: zero third-party services and no network calls — nothing leaves the machine, so it runs air-gapped and sidesteps data-egress / residency review; deterministic and fully auditable (readable Python, no black-box model) for security sign-off; built-in secret redaction (API keys/tokens masked before they ever reach context, logs, or the cache); an owner-only (0700) local cache; reversible (originals retained); open-MCP-standard integration that drops into an approved toolchain; and MIT licensing for legal clearance.