sondalab

Sondalab · a lab for operators

Wired for operators,not passengers.

see it · steer it · own it

See what we wireWhat we believe →
01The stance

Tools that keep you in command of the machines working for you.

Sondalab builds software that runs, watches and improves the agents doing your work, and keeps them legible, controllable and yours. It's a stance that holds as machines grow more capable: from today's software agents toward the physical systems, and robotics, ahead.

  1. See it

    Whatever a machine is doing, Sondalab keeps it legible: observable, measurable, understandable as it happens. No black boxes.

  2. Steer it

    You stay in the loop and in charge: direct it, correct it, hand off or take the controls back at any moment. Assistance, never autopilot.

  3. Own it

    The machine, its data and its off-switch stay yours, answering to you, running where you are, not to a service out of reach.

02The tools

Many tools, one throughline.

Together they cover the whole loop of working with agents: writing and running the work, watching every step as it happens, scoping what a session loads, hosting it all on hardware you own, and extending what an agent can do. Different forms (an IDE, dashboards, a launcher, a runtime, a toolbox), one stance underneath: local-first, legible, and yours to switch off.

  1. 1.

    SPEXR

    Public beta

    The spec-based IDE where agents do the work and you keep command.

    An IDE on Eclipse Theia where the agent session is the main surface and every change traces back to a spec — plan, tasks, diff, PR. With on-device semantic search and Darkfactory, a live wall of your agent sessions.

    IDE
  2. 2.

    introspect

    Beta

    Audit and observe how your coding agent runs, replay and live.

    A local, read-only dashboard over the files your coding agent already writes — instructions, skills, memories, hooks, permissions — with past sessions replayed as execution graphs. It never writes to your config.

    App
  3. 3.

    claude-loadout

    Start each session with only the tools it needs.

    Ranks your installed MCP servers, plugins, skills and notes against what you're working on, with a local model, then starts the session with only the relevant subset — the rest installed, just unloaded.

    CLI
  4. 4.

    local-agents

    Your agents, on your machine.

    A router in front of LM Studio that sends each request to the right local specialist — coder, creative, home assistant — with voice and a dashboard to assign models, all on your own GPU.

    App
  5. 5.

    skillkit

    A pick-and-install kit of agent skills.

    Five skills that give a coding agent durable working habits: decisions with a revert plan, specs and memories the repo can see, the set synced across machines. Two of the five now ride inside claude-loadout too.

    Kit

Research

Where we test an idea before it becomes a tool: whether an answer can be made to repeat, and whether a coding agent can carry its project's experience from one session to the next.

  1. 1.

    daos

    A research bench for getting the same answer twice.

    A Python library we built to make an AI model give the same answer every time, and to measure what that takes. The findings are in the research note.

    Library
  2. 2.

    MxM

    Your coding agent remembers every session, on your machine, where you can read it.

    Stores what past coding-agent sessions said and feeds the next one its project's experience, recalled by meaning with a local encoder, in one SQLite file per repository that you can open, move or delete.

    CLI
Public beta
usable, still stabilising
Beta
early, evolving fast
03The method

We run on our own tools first.

Sondalab is built by directing agents with the same tools we ship. What reaches you has already carried our own work.

  1. Intent

    We write the intent first.

    Every change starts as a written decision, before an agent touches a file.

  2. Work

    The machines show their work.

    Each step is visible while it runs and traceable afterwards, so nothing lands unexplained.

  3. Check

    Drift doesn't ship.

    A design audit and a content schema stop the build on any mismatch. This page passed both.

04Contact

Building something you'd rather run yourself?

Sondalab is early and independent. If the operator's stance resonates, as a user, a builder, or a collaborator, we'd like to hear from you.