LOCI v0.1 (Date: 02 Jul 2026)

Coding agents write code. LOCI thinks ahead: it predicts what AI-generated code will do — time, power, memory, and system behavior — during planning and code writing, before any of it runs. Now directly inside Claude Code, at exactly those two moments: Plan Guard while Claude drafts a plan, Save Guard the instant it writes.

New

Two stations run automatically, the moment Claude plans or writes:

• Plan Guard (loci-preflight) — Before Claude writes a line of code, LOCI audits the plan at the binary level — timing, power, and control-flow impact of the functions involved — and returns a verdict: GOOD, ADJUST PLAN, or STOP.

• Save Guard (loci-post-edit) — Every source edit is compiled and diffed against the pre-edit binary automatically. LOCI returns a per-function regression verdict — OK, CAUTION, or FLAG — and proposes a fix when it flags one.

Five more are yours to call on directly, in plain language:

• exec-trace — Ask "What's the execution cost of main()?" and get function-level timing and power, grounded in real workloads and platform traces, with a confidence band on every number.

• stack-depth — Ask "Is my stack safe for TaskMain?" and get worst-case stack depth along the deepest call path, checked against your budget, with RTOS task entry points detected automatically (xTaskCreate, Task_construct, osThreadNew, FreeRTOSConfig.h).

• memory-report — Ask "How much ROM/RAM does my build use?" and get a section-by-section breakdown of the compiled binary and its top consumers.

• control-flow — Ask "Show me the call graph for process_data()." and get an annotated control-flow graph — every branch, every call edge — with indirect-call hazards flagged by name.

• trends — Ask "How are my functions doing?" and get per-function timing, power, stack, and memory history on the current branch, broken down by improved, regressed, and baseline.

Plus three that keep the other five running smoothly:

• /help — Shows your environment status, sign-in state, and the full skill list, with a contextual next step.

• /loci:setup — Reinstalls and repairs the loci CLI. Idempotent — safe to run any time something looks off.

• /bug-report — Generates a full diagnostic report when a skill doesn't fire the way you expect.

And the record of all of it:

• LOCI Portal — Every run from Claude Code is logged automatically, with a Sessions log, configurable gate thresholds, and PR-level review against your open pull requests.

• Session hooks — SessionStart detects your project and build environment automatically; PreToolUse snapshots the binary ahead of every edit so Save Guard has a baseline to diff against.

Good to know

• Every result comes from the compiled binary directly — no profilers, no runtime instrumentation, no connected board, and no source code changes.

• The gate is advisory, not automatic: LOCI proposes a verdict, a human makes the final call.

• stack-depth, memory-report, and control-flow run entirely from the local binary — no MCP connection or network dependency required.

Prerequisites

Requirement	Version	Needed for
Claude Code	Latest	Everything
Python	3.12+	The loci CLI (local ELF/build analysis)
uv	Any	Installing the loci CLI
jq	Any	Session hooks
Compiled binaries (.elf / .o / .axf)	—	All skills
Network access to the LOCI backend	—	Plan Guard, Save Guard, on-demand timing/power analysis

One cross-compiler, matched to the target: arm-none-eabi-gcc (ARM Cortex-M), aarch64-linux-gnu-gcc (ARM Cortex-A), tricore-elf-gcc (TriCore), tiarmclang or armcl (TI ARM), g++ or clang++ (x86/x64). stack-depth, memory-report, and control-flow run without a cross-compiler or MCP connection at all.

---

LOCI works alongside Claude Cod. This is our first release for Claude Code specifically — more AI Agents Cursor, Copilot, and your own agents; deeper CI/CD integration; and broader toolchain coverage are next.

Questions or feedback? [e-mail]