Patent pending · U.S. Provisional Application No. 64/082,906

Coordinate · Check · Repair · Gate

Conduct the effort.
Relay the result.

Conductor Relay coordinates AI work across your model team, checks, repairs, and approval gates — turning raw model output into governed output your team knows how to handle.

local-firstworks with your model terminal-readyreview before you use

Save money running local or hybrid — see the cost model calculator.

Watch CR Lite check, repair, label, and gate a model’s output — 49 seconds.

The problem

Running local AI is easy.
Trusting the answer is the hard part.

Tools like Ollama and LM Studio can run a model on your computer. But they do not tell you whether the answer is correct enough to use, safe to save, or ready to send somewhere else.

01

AI output can look finished and still be wrong

A model can give you code, JSON, reports, or plans that look good at first glance. But the output may have broken syntax, missing fields, fake claims, hidden mistakes, or unsafe content.

02

A failed check is not enough

A normal validator can tell you something failed. CR Lite goes further. It checks the output, gives the model clear repair feedback, checks again, and shows you the final result.

03

The model should not grade itself

The model can write the answer. A second model can review it. But CR Lite's checks and gates decide the label. The model does not decide what is trusted, saved, exported, committed, deployed, or acted on.

How it works

Every answer goes through the gate.

CR Lite sits on top of your runtime — bring your own model. It takes whatever the model produced and puts it through a clear pipeline before you ever rely on it.

Label — how far to trust an output
CR Lite’s plain verdict on each answer. VERIFIED passed the needed checks · UNVERIFIED_REVIEW usable, review first · NEEDS_REVIEW something important still needs checking · BEST_EFFORT best available, not safe to trust without review.
Gate — what you can do with it next
The rule that allows or holds the next step — save, export, run, commit, deploy, or act — based on the label. Weak or unproven output can’t move forward unreviewed; you stay in control of every decision.
Step 1

Turn the request into a task

CR Lite turns your prompt into a clear task so the model cannot silently change what it is supposed to do.

Step 2

Generate the first answer

Your model creates the first output.

Step 3

Check the output

CR Lite checks the output for things like syntax, format, missing requirements, source needs, secrets, and basic runtime problems.

Step 4

Repair what failed

When something fails, CR Lite gives the model clear feedback and asks for a better version.

Step 5

Label the result

CR Lite gives the final result a label:
VERIFIED passed the needed checks
UNVERIFIED_REVIEW usable for review, not fully proven
NEEDS_REVIEW something important still needs checking
BEST_EFFORT best result available, not safe to trust without review

Step 6

You decide what happens next

CR Lite shows the answer and the label. You decide whether to use it, review it, save it, export it, commit it, deploy it, or try again.

filesshellrepo search git diffcommit guardsyntax testssecretssource checks skillsmemoryexpandable tools

Checks find the problems. Repairs fix what they can. Confidence labels the result. You always see the best available output.

Capabilities

Tools and skills built in.

CR Lite is not just a chat box. It is a local AI workbench with tools, skills, memory, Git safety, and repair loops around the model.

The model can ask for help. CR Lite decides what is safe to run.

Bash
Runs shell commands in a controlled way.
Read
Reads files, images, PDFs, notebooks, and project text.
Edit
Makes exact edits to existing files.
Write
Creates or overwrites files.
Agent
Sends work to a helper agent.
Workflow
Runs a planned multi-step workflow when you choose it.
Skill
Uses a saved skill for a known job.
Tool Search
Loads tool details when a tool is needed.
Ask User
Asks you a clear multiple-choice question.
Schedule Wakeup
Lets long loops pause and continue later.

Tool calling, but governed

A normal AI tool call can be risky. It may read the wrong file, write to the wrong place, run the wrong command, or act before you approve it. CR Lite adds rules around tool use.

Safe read-only tools can run when needed. Risky actions like writes, commits, deploys, money movement, secrets, or production changes require approval.

The model can ask. CR Lite checks. You approve the risky parts.

Start simple. Expand when you need more.

Expandable tools optional · add when you need them
web searchweb page readingnotebook editing design syncbackground task monitoringpush notifications scheduled jobstask creation and trackingremote triggers project workflowsrepo searchcode review tools security review toolsverification tools

These tools provide evidence. They do not decide what is trusted.

Git-aware by design

AI coding gets risky when it touches real files. CR Lite can help you work with Git safely. It can inspect changes, explain diffs, check what files were touched, and help prepare a clean commit — without letting the model silently push or mutate your repo.

Status check
Shows what files changed before anything is saved or committed.
Diff review
Explains what changed and why it matters.
Scoped staging
Adds only approved files, never a broad git add -A.
Commit guard
Blocks unsafe paths, secrets, private evidence, and out-of-scope files.
Commit message help
Drafts a clear commit message based on the actual change.
Branch awareness
Warns before risky commits on main or master.
Worktree support
Keeps larger work isolated when needed.
PR prep
Helps summarize what changed for review.
The model can suggest a Git action.
CR Lite checks whether that action is safe.
You approve the risky parts.

CR Lite lets AI help with code and Git, but keeps repo changes gated, scoped, and reviewable.

Skills for repeatable work

Skills are saved ways of doing common jobs. Instead of asking the model to figure out the process every time, CR Lite can use a skill with clear steps, rules, and checks.

Built-in process skills click to expand
brainstormingwriting plansexecuting plans systematic debuggingtest-driven developmentcode review security reviewverification before completionsub-agent development parallel agent workgit worktree workflowsfinishing a development branch writing new skills
Utility skills click to expand
reviewsimplifyverifyrun initialize a projectdeep researchAPI help loop managementschedulingconfig updates keybinding helpstatus line setup

Without skills, every prompt starts from scratch. With skills, CR Lite can follow a known process:

  1. understand the request
  2. make a plan
  3. run checks
  4. repair problems
  5. verify the result
  6. show you what is safe to use
Skills make the model less random and the work more repeatable.

Scoped LLM memory

CR Lite can use memory, but it does not dump all memory into every prompt. Each role gets only what it needs.

  • the writer gets task and style context
  • the validator gets review context
  • the tool planner gets tool-use context
  • private or secret data stays out of prompts
This helps the model stay consistent without giving it unchecked power.
The proof

The writer is not the checker.

The model writes the answer. A validator model may give advice. But CR Lite's checks and gates decide the label — and what you can safely do next.

WHO DOES WHAT
Writer model
writes the answer
Validator model
may give advice · a separate model
CR Lite checks
decide the label · not a model
That separation matters. The model never gets to say, "Trust me."

CR Lite shows you what passed, what failed, what was repaired, and what still needs review. You decide what gets trusted, saved, exported, committed, deployed, or acted on.

Model roles

Use one model or many.

CR Lite works with any model. Use one model for every role, or give each role its own — local, an API, or a mix. No matter what you connect, the label is set by CR Lite's own checks — fixed rules and tests, not another AI grading the work.

ONE MODEL

One model, every role

A single local model fills every role. The simplest way to start.

TWO MODELS

One writes, one checks

One model writes the answer; a second model reviews it.

FOUR ROLES

A model per role

Give Report, Code, Tool-call, and Validator each their own model.

MIXED STACK

Local + API together

Run a local model and add a stronger API checker when you need it.

One model or four different ones — the label still comes from CR Lite's checks (fixed rules and tests), not from an AI judging the answer.
The CLI workbench

One terminal. Any model.

Connect one model to every role, or wire different models to Report, Code, Tool-call, and Validator. Either way, the label comes from CR Lite's checks — fixed rules and tests, not another AI grading the work.

Local-first: when you use local models, work runs on-device. The result still carries a confidence label and a gate before save, export, or action.

Advanced loops

Built to improve without losing control.

REPAIR

Repair loops

When an answer fails, CR Lite can give the model clear feedback and ask for a better version. It does not just say "wrong." It says what failed and what needs to change.

MEMORY

Memory loops

CR Lite can keep useful project context so future runs do not start from zero. Memory is scoped. The model only receives the context needed for that role and task.

INDUCTION

Induction loops

For advanced use, CR Lite can use past failures and successful repairs to improve future checks and workflows.

The loop can improve the process, but CR Lite still checks the result before anything is trusted, saved, exported, committed, deployed, or acted on.
Free hosted demo

Prospects get evidence. Customers get CR Lite.

The free hosted demo uses controlled scenarios only. It shows the CR Lite workflow without exposing the engine, accepting private repo data, or running arbitrary customer code.

Hosted demo — free.
Controlled scenarios. Evidence of the product and engine. No download, no private code, no arbitrary execution.
Paid local license.
The paid local license runs the CR Lite engine on your machine. Your local workflows, repos, and model outputs stay local unless you choose to connect external tools under your own configuration.

Stop guessing what AI output is safe to use.

CR Lite checks, repairs, and labels local AI output before you rely on it.

The model proposes. You decide.