Skip to content
Kronumos Core v1.3.0 · Rust CLI & MCP Harness

Terminal AI coding agent. Fewer turns, lower token spend.

A fast CLI coding assistant and autonomous debugging harness in Rust. Reduces LLM token costs by 91.3%, validates multi-file patches with Tree-sitter AST before disk write, and verifies every fix on your real test suite.

1,452 framework frames pruned in 0.4ms4,820 tok
Select Failure ScenarioAuto-Play
$
◆Probe workspace baseline -> 1 of 18 assertions failed
◆Redacted postgres://[REDACTED] locally in Rust memory
◆Pinpointed failing line at src/auth/service.py:41
1,480 raw lines28 clean lines
01 SHIELD0.01ms
02 SURGERY-91.3%
03 AST GUARD0.2ms
04 VERIFYPASS
Filtered locally before LLM promptRust Sub-Cortex4,820 tok
- node_modules/next/dist/compiled/webpack/bundle5.js:29482
- site-packages/starlette/_exception_handler.py:53
- DATABASE_URL=postgres://admin:[email protected]/prod
src/auth/service.pyScanning AST...
39  def verify_session_token(claims: dict) -> bool:
40      exp = claims.get("exp", 0)
41 - return exp > current_epoch
41 + return int(exp) > int(time.time())
42  
43  def revoke_token(token_id: str) -> None:
pytest tests/test_auth.py -> 18/18 passed (410ms · 0 dirty diff)Try in Web Agent →
Works with Kronumos Cloud · OpenAI · Gemini · Claude (OpenRouter) · Local OllamaSupports Rust · Python · TypeScript · Go · C++
Chapter 01 · The Context Tax

Why coding agents burn your token quota before fixing one bug.

Most AI coding assistants stream unfiltered stack traces directly from failing tests into the reasoning model. Thousands of library lines drown out the single file you care about, forcing repeated retry loops when syntax breaks on disk.

Raw Noise
1,480 lines
Retry Loops
3 to 6 turns
Kronumos
1 clean pass
Token Variation Per SWE-bench TaskMeasured on 500 tasks
Standard Unfiltered Agent Loop38,412 tokens / task
Raw stack traces + syntax retries100% spend
Kronumos AI Coding Agent Harness2,512 tokens / task
-91.3% overhead pruned locally in Rust
Context Discipline
Only failing workspace lines sent
Zero Retry Churn
Broken syntax blocked pre-write
Quota Longevity
More tasks per API dollar or plan
Chapter 02 · How Kronumos Solves It

Four deterministic gates inside an AI coding agent in Rust.

Before a single prompt hits the model and before any patch touches your git tree, every step runs through a native Rust pipeline in under a millisecond.

Stage 01 · 0.008ms · local Rust

Masks credentials in memory before any prompt leaves your machine.

Environment variables, API keys, JWTs, and database connection strings are redacted inside the native Rust Sub-Cortex before touching the network.

Read Technical Docs
Raw Workspace Context
- STRIPE_SECRET_KEY=sk_live_98f2a4c...
- DATABASE_URL=postgres://admin:[email protected]:5432/app
- Authorization: Bearer eyJhbGciOiJIUzI1NiIs...
Sanitized Prompt Payload
+ STRIPE_SECRET_KEY=[REDACTED_KEY]
+ DATABASE_URL=postgres://[REDACTED]@db.internal:5432/app
+ Authorization: Bearer [REDACTED_JWT]
Chapter 03 · Production Proof

Engineered for real terminals, MCP tooling, and automated test repair.

Inspect AI coding agent benchmarks
01 · Slash Command Palette & MCP/blast · /hunt · /mcp
kronumos > /AST ScopeGuard Ready
/blastTrace cross-file impact & symbol dependents
/huntStatic audit for edge cases & unsafe unwraps
/mcpAttach Model Context Protocol tools locally
/setupSwitch Cloud, OpenRouter BYOK, or Ollama
/usageInspect 5h rolling & daily quota pools

One terminal workflow from impact check to patch verify.

Inspect dependency graphs, monitor token pools, and invoke MCP servers without leaving your keyboard.

02 · Adversarial Bug Hunt & Test GateAutomated Repair Loop
$ kronumos /hunt src/metrics.py<1ms local AST
FindingZeroDivisionError in compute_success_rate
Pre-Write CheckTree-sitter + ScopeGuard PASS
- return passed / total
+ return passed / total if total > 0 else 0.0
pytest tests/test_metrics.py14/14 passed · 0 dirty diff

Fixes verified against real test suites, never guesswork.

Audit subtle edge cases locally and verify every diff against your real project tests before anything stays in git.

Tokens Per Task
2,512
Down from 38,412 unfiltered
Clean Git Apply
442 / 442
100% AST-healed patches
Local Pre-Write Check
<1ms
Rust Sub-Cortex & ScopeGuard
SWE-bench Verified
500 / 500
Bare-metal evaluation suite
Chapter 04 · Run Anywhere

Terminal CLI, local LLM support, and independent systems research.

Cross-Platform & Headless CI

Static Rust binary runs natively on Windows PowerShell without WSL, macOS Apple Silicon, and automated Linux CI/CD pipelines.

Setup for Windows, macOS & Linux

OpenRouter BYOK & Local Ollama

Never locked into one vendor. Start on Kronumos Cloud, connect OpenAI, Gemini, or Claude via OpenRouter, or run air-gapped with Ollama local LLMs.

View air-gapped & team options

Research at Tokenectomy Labs

Kronumos translates peer-reviewed compiler verification and AST surgery research from Tokenectomy Labs directly into sub-millisecond developer tooling.

Explore labs.kronumos.com
FAQ

Frequently asked questions

Most AI coding agents stream raw stack traces and third-party dependency frames straight into the LLM, burning context and repeating broken syntax loops. Kronumos runs a local Rust Sub-Cortex harness that prunes 91.3% of stack-trace noise in under a millisecond and blocks broken syntax pre-write, cutting median task consumption from 38,412 tokens to 2,512 tokens.

Start now on your repository.

Launch in your browser or run natively in your local terminal.