Sparse is a token-efficient AI coding agent that strips waste before the model is ever called. You set a hard token budget up front — no surprise bills, no runaway loops.
The failure is in planner.ts:84. When the dependency graph cycles, visited isn't reset before the second pass. Proposed patch:
Modern coding agents shove entire files, terminal logs and old chat history into every model call. That drives cost up, slows iteration down, and leaves you with unpredictable bills that can hit hundreds of dollars a month per developer. Research shows trimming irrelevant context actually improves task success — waste is the enemy, not context.
Agents pack irrelevant files and stale history into every request, drowning the model in noise and burning tokens on zero-value context.
Long context windows sound powerful — until your monthly bill lands. Agent sessions can run hundreds of dollars per developer per month, and most of those tokens do nothing.
You can't see what the agent sent, or why. Without hard budget caps, you don't control the cost before the run starts — you just get the bill after.
Sparse uses repo maps, hard token budgets, log compression and a context firewall to strip unnecessary tokens before the model is ever called. You decide what a run may cost before it starts.
A lightweight symbolic index of your codebase — imports, exports, call graphs — instead of raw files.
Choose Cheap, Balanced or Deep. You decide what a run may cost before it starts, and Sparse enforces that cap — no runaway loops, no surprise bills.
Terminal output and stack traces are distilled to the lines that actually explain the failure.
Rules that block secrets, node_modules, generated files and stale chat from ever reaching the model.
Every feature is designed around one goal: send less, get better results.
Sparse picks the minimum set of files, symbols and diffs the task actually needs.
Trade cost for depth on the fly. Hard budget caps per run mean you control the spend before the agent starts, not after the bill arrives.
Your source stays on your machine. Sparse reads it from disk — not from a hosted database.
Bring your own Anthropic, OpenAI or open-source model keys. You control the spend and the vendor.
Sparse proposes structured patches you can review and apply — not free-form file rewrites.
Every run shows what was included, what was compressed, and how many tokens you saved.
Point Sparse at any local git repo. No upload, no sync.
A lightweight, symbolic index of your code — updated as you work.
Ask anything. Sparse assembles just the files and logs the task requires.
You bring the API keys. We make every one of them count — with hard budget caps and a token receipt to prove it.
Everything you need to try Sparse.
For developers shipping every day.
Pro costs less than the tokens it saves you — if you spend $40+/mo on API keys, it pays for itself.
Early-access founders: $9/mo locked forever — assigned manually to waitlist users.
Shared configs, roles, billing.
Sparse is designed as a desktop client. The web app handles your account and authentication — nothing else. Source code, prompts, terminal logs and API keys never touch our database.
Free during early access. BYO API keys. Set your budget before the run starts. Windows, macOS and Linux desktop apps coming soon.