Private beta Sidekick is taking applications

Lead an AI engineering team.

Sidekick is not a tool you run. It's a team that works with you — planners, builders, reviewers, an evaluator. You steer, it does the engineering, on your own box and your own forge, with every dollar accounted for. Pull the cable, the team keeps working.

For a small group of solo developers and small teams. Setup help, a direct line to the founder, and a say in what we build next.

sidekick · session · sandboxed armed
planplans/your-feature.mdloaded builderbranch pusheddone reviewerchanges requested · resolveddone evaluatorjudged against the planpassed gatetouches tests/ · protected pathwaiting on you

transcript stored · redacted

Illustration of a session. Protected paths always wait for a human.

How it works

From an idea to production, in three phases.

  1. 01

    Plan

    Start in the chat you already use. The team plans with you, not for you — you write the contract, it signs the work.

  2. 02Optional phase Optional

    Design

    Artwork, tokens, contracts, fixtures, drafted as candidates. You pick. Nothing builds until you sign off.

  3. 03

    Build

    One pull request per step, reviewed and judged against the plan. You're asked once, not on every step.

    1. You decide1 ContractYou write the plan, a PRP. It's the contract.Sign-off: what gets made
    2. Automated2 BuildBuilders work it in sandboxes.
    3. Automated3 ReviewReviewers check every change.
    4. Automated4 EvaluateAn evaluator judges against the plan.
    5. You decide5 MergeArmed, Sidekick merges. Protected paths wait for you.Sign-off: what ships

    Fix loop: Review or Evaluate sends the work back to Build.

    • Unarmed Merge waits for you.
    • Armed Sidekick merges. Protected paths still wait for you, and the receipt records what it passed.

Control

You stay in command.

The final say is yours

Unarmed, every merge waits for you. Armed, Sidekick merges and states what it passed. Protected paths always wait for a human.

Every task, job and run has a receipt.

Ask for a review, a feature, a bug fix, or a full run. Each lands with gate facts, judge evidence, fix rounds, and a real dollar figure. Nothing priced as free.

Rate limits are a fallback, not a stop.

When the subscription lane hits its ceiling, the run switches lanes and keeps working.

Unattended, never unseen

Redacted transcript for every session. Blocked work is saved, not lost.

Off the grid

Runs on your own box. Internet optional.

Pull the cable and the team keeps shipping. Forge, models, agents and search all live on your machine — and your cloud models work through the same gateway. Open weights or closed, one gateway, no vendor owns the stack.

your box offgrid
networkunpluggedoff forgepull requests, review, mergeup llmslocal models, one gatewayup agentssandboxed builders and reviewersup ragcode and docs, indexed locallyup
Open-source release coming

open-llm-stack

The foundation underneath Sidekick. One gateway in front of your local models and any provider, with a budget on every key and a switch that takes you fully offgrid.

Private by default
Request bodies aren't stored unless you turn logging on, and then only through a redaction hook.
quickstart
$ make init
$ make up
$ make test
# zero edits to get started

Every superhero needs a sidekick they trust.

The private beta is open to solo developers and small teams with real work to ship.

Apply for the private beta