Introduction
Gremlin is an AI developer that keeps working after you close the laptop. It replaces your day-to-day Claude Code with one conversation, on your desktop and in your terminal, and runs a sandboxed developer around the clock.
Local models do the cheap, high-volume work: drafting, testing and checking their own output. Claude plans the task and gives the final review. Every task ends as something you can read in a minute: a brief, a diff, the checks that ran and what it cost. Nothing is merged, deployed or posted until you approve it, from the desktop or from your phone.
Already have a .claude/ folder? Gremlin reads it as it is. Your skills,
agents and settings come with you.
How it works
Each step goes to the cheapest model that can do it. Each role's model can be changed in settings; these are the defaults.
| Step | Local modelThe gremlin | ClaudeThrough the official CLI | YouOne tap |
|---|---|---|---|
| PlanBrief and approach | — | Yesa call or two | Answers questions |
| DraftCode, docs, fixes | Yesas many tries as it takes | — | — |
| TestCompile, lint, tests | Yesfailures loop back here | — | — |
| Self-checkReview its own diff | Yes | — | — |
| Final reviewBefore it reaches you | — | Yesonce, at the end | — |
| ShipMerge, deploy, post | — | — | Approvedesktop or phone |
> fix the flaky login test and open a PR
Planning with Claude (brief, 1 call)
Brief ready. Handing to the local model.
Tinkering... edited 2 files. Running tests.
Tests pass. Lint clean. Local review clean.
Claude final review: looks good, one nit fixed.
Needs you: open pull request? [approve] [reject]
An illustrative session, not a recording.1
While you sleep
Every task is a saved state machine. A crash, a reboot or a usage limit only pauses it; it picks up from the last checkpoint. Gremlin spends Claude last, can spread work across several Claude accounts, and keeps the queue moving on local models while a usage window resets.
One night, four tasks
IllustrationClaude appears twice per task: once to plan, once to review. Everything in between is local.
Work done over a day
IllustrationGremlin picks an account that has room, or waits on the local model. No task is dropped.
Approve from the sofa
IllustrationNeeds you
Open PR: fix flaky login test
3 checks passed. Reviewed by Claude. Cost $0.04.
A real app for iOS and Android. Read the brief, the diff and the cost, then tap. High-risk approvals need your fingerprint or face. The link to your desktop is outbound-only, so you never open a port at home.
Safety
Default-deny sandbox. The agent runs where it can reach only the model endpoints and hosts you allow. No network, no secrets and no surprises until you open a door.
Keys stay home. Credentials live in your vault. The sandbox holds only a short-lived token for its own task, and never sees an API key or a password.
Approval first. Merges, deploys, external posts and ticket changes wait for a person, unless you wrote a per-project rule that says otherwise. A stop-all switch halts everything at once.
An audit trail. See what ran, which model did it, what it cost and who approved it.
Works with
| Desktop | macOS (Apple Silicon), Windows 10 and 11, Linux (Arch and Debian/Ubuntu) |
|---|---|
| Phone | iOS and Android |
| Local models | Any OpenAI- or Anthropic-compatible endpoint you add, on any hardware |
| Claude |
The official claude CLI, with your subscription login or an API
key; several accounts at once
|
| Trackers | GitHub, GitLab, Jira, Linear |
| Extensions | MCP servers, Agent Skills, Claude Code plugins, notifiers |
Small core; everything else is an extension on one SDK.
Get started
Put a gremlin to work.
Installers are not released yet. Gremlin runs on your machine, with no telemetry and no analytics; your code goes only to the endpoints you configured.2