Ctrlk
AuroraLabsGuardian for coding agentsGround Your Changes in Execution Evidence
For the complete documentation index, see llms.txt. This page is also available as Markdown.

LOCI Plugin for AI Coding Assistants

AI Writes Code. LOCI Gates It.

LOCI is a quality-gate agent for Claude Code that models regressions, power, latency, and control-flow safety directly from your compiled binary, from plan to merge. It requires no running code, no instrumentation, and no code changes: every result is grounded in real execution data recovered from the binary itself.

What LOCI does

LOCI connects Claude Code to LCLM (Large Code Language Model), a model trained on billions of assembly blocks and real hardware traces from IoT, networking, and safety-critical systems — not a general-purpose model wrapped around your code. Because analysis happens at the binary level, LOCI can answer questions about timing, energy, memory, and control flow that source-level tools can't: the actual compiled output, not an approximation of it.

Prerequisites

Requirement

Version

Required for

Claude Code

latest

Everything

Python

3.12+

The loci CLI (local ELF/ build analysis)

uv

any

Installs the loci CLI as a tool. Install this yourself — the plugin checks for it and gives you the install command if it's missing.

jq

any

Session hooks. Install this yourself - the plugin checks for it and gives you the install command if it's missing.

Cross-compiler depends on the project you work with.

Skills that work without a cross-compiler or backend network access: stack-depth, memory-report, control-flow.

Install

Install the plugin from inside Claude Code:

/plugin marketplace add https://github.com/auroralabs-loci/loci-claude

/plugin install loci@loci

/reload-plugins

/loci:setup

There is no separate CLI download step. The loci command-line tool is installed automatically in the background the first time a Claude Code session starts, using uv. If uv or jq aren't already on your system, Claude will tell you the install command to run.

Authentication

Timing and energy analysis reach the LOCI backend and require a signed-in session.

  1. Run ! loci login from inside Claude Code (or loci login in your terminal).

  2. Confirm it worked with loci auth status - it should show signed_in.

  • stack-depth, memory-report, and control-flow run fully offline.

  • exec-trace, loci-preflight, and loci-post-edit require a signed-in session because they call the LOCI backend.

Quick Start

Once installed, try these in any C/C++/Rust project with compiled binaries:

  1. Timing & energy — ask: "What's the execution cost of main()?"

  2. Memory budget — ask: "How much ROM/RAM does my build use?"

  3. Stack safety — ask: "Is my stack safe for TaskMain?"

  4. Control-flow safety — ask: "What does the call graph for process_data() look like?"

Two skills also run automatically, with no invocation needed:

  • loci-preflight fires during /plan mode — analyzes callees at the binary level before code is written.

  • loci-post-edit fires after every edit to C/C++/Rust source — diffs the binary and returns a regression verdict.

Skills Reference

Skill

Trigger

Sign-in required

What it does

loci-preflight

Automatic, in /plan mode

Yes

Audits the plan at binary level before code is written — timing, energy, and control-flow-graph impact.

loci-post-edit

Automatic, after edits to C/C++/Rust

Yes

Diffs pre/post compiled artifacts and returns a regression verdict on timing, energy, and control flow.

exec-trace

User-invoked

Yes

Function-level timing and energy from real-time hardware traces, powered by LCLM.

stack-depth

User-invoked

Yes

Worst-case stack depth via call-graph traversal and per-function frame sizes.

memory-report

User-invoked

Yes

ROM/RAM section breakdown and top consumers from compiled ELF binaries.

control-flow

User-invoked

Yes

Annotated control-flow graphs, optimized for LLM analysis.

trends

User-invoked

Yes

Per-function measurement history and optimization progress on the current branch.

help

/help

No

Confirms LOCI is loaded and lists available skills.

setup

/loci:setup, or automatically on CLI-missing errors

No

Installs, repairs, or verifies the loci CLI. Safe to re-run at any time.

bug-report

/bug-report

No

Generates a full diagnostic report when something isn't working.

Automatic Hooks

The plugin registers four hooks that run without user action:

Hook

Trigger

What it does

SessionStart

Session startup

Detects your project's build system, compiler, and target architecture; installs the loci CLI in the background if it's missing; and injects session context (auto-run rules, tool policy, response tone) for Claude to use. Never blocks session startup, even if detection or install fails.

PreToolUse

Before Edit, Write, or MultiEdit

Snapshots the current compiled .o for later delta analysis and runs a static call-graph pre-scan of the incoming code (before it's written to disk), surfacing any findings as context. Advisory only — always exits 0, never blocks the edit.

PostToolUse

After Edit, Write, or MultiEdit

Checks whether the change could alter a compiled function's body and, if so, injects a reminder instructing Claude to run the loci-post-edit skill immediately. Advisory only — always exits 0, never blocks.

Stop

End of session

If the loci CLI is present, flushes any pending analysis results to your account (loci stats flush-impacts).

Troubleshooting

loci CLI not found:

  1. Check it's on PATH: command -v loci and loci --version.

  2. If it's missing, the background install may still be running or may have failed. Run /loci:setup to reinstall and verify — it's safe to re-run — or loci doctor once the CLI is present.

  3. Confirm uv is installed (command -v uv); the plugin needs it to install the CLI.

A skill reports auth_required:

  1. Run ! loci login in your terminal, then retry.

  2. Confirm loci auth status shows signed_in.

Auto-skills didn't fire:

  • loci-preflight only runs in /plan mode — confirm you're planning new logic, not just asking a question.

  • loci-post-edit only runs after edits to C/C++/Rust source files.

  • Both compile the source themselves — they don't need a pre-built binary. They do require a supported LOCI target and a working cross-compiler for it on PATH, plus a valid session context from startup. If the compiler is missing, restore it (or point LOCI at it) and re-run.

On-demand skills didn't respond:

  • Type /help to confirm LOCI is loaded and see the full skill list.

  • Restart Claude Code from inside your project directory if the build environment wasn't detected at session start.

  • Check that a cross-compiler is installed and on your PATH (see the table above).

Nothing seems to work:

Run /bug-report to generate a full diagnostic report.

LOCI Portal

Every analysis run from Claude Code is logged and visualized automatically at the LOCI Portal (app.auroralabs.com), no extra configuration needed. The Portal is where you view session history, binary analysis results, PR review status, and team/quota settings.

Account Plans

LOCI is available on a free tier with limited daily interactions. Paid tiers raise or remove that limit and unlock additional capabilities...

For current plans and pricing, see auroralabs.com/pricing.

Support

Reference

  • Quick Start (web): auroralabs.com/product#quick-start

  • Talk to us: auroralabs.com/contact

  • LOCI install endpoint: https://loci.auroralabs.com/install.sh

PreviousWhich path is right for you?
NextHow LOCI fits your loop.

Last updated

Was this helpful?