Demos

Real VMs, not just containers

Each branch runs in its own Linux VM, so ports, databases, queues and sessions can't collide and the host stays clean. VMs are copy-on-write snapshots of a baked base: 0 MB and under a second for an 8.6 GB image.

Built for agents, with MCP inside

Every command has JSON output and exit codes an agent can act on. When a VM is up, the MCP servers inside it are registered with the agent automatically, so the agent gets the platform's tools as well as a shell.

Mocks the app can't see

A mock answers on the providers' real hostnames using a throwaway certificate authority, with no changes to application code. Anything that must reach the real internet goes out through a proxy on the laptop, which returns each reply to the VM that asked.

Careful deletes

Cleanup checks again for uncommitted or unpushed work at the moment it deletes, and reports anything it won't touch.

Boot and agent steps are sped up here. The memory limits are the real ones.

Small, explicit MCP tools

Agents get one tool per action instead of a save-everything endpoint, so a stale copy can't overwrite newer work.

One path for everyone

The product's own interface and the agents use the same endpoints, so the same rules protect both.

Guardrails on every write

Validation runs before anything is saved, so a bad request comes back as an error that names the field instead of becoming bad data.

A person stays in charge

Agent changes are flagged in the history, editing an approved plan sends it back for review, and reviewers are told.

Annotate, don't rewrite

A route is exposed to MCP by annotating it where it already lives. Nothing is duplicated into a hand-written MCP server.

Generated, so always in step

Tool names, argument schemas and descriptions come from the API spec, and the spec is checked in CI so it can't drift from the code.

Deployed with the spec

Each MCP server only starts with the spec it was built for, after new tools once went missing until a restart.

Agents can point, not just tell

A shared route-map tool lets any agent hand you a link to the exact page, instead of describing where to click.

Hours, not seconds

Expansion, classification and metrics run as background jobs grouped into a batch, with each chunk its own task on the worker pool. Task time limits had to grow from one hour to six, and a watchdog restarts a worker that stalls, so one stuck chunk never takes the batch down with it. The biggest batches run for the best part of a day.

Chunks and pages everywhere

Classification goes to the model in chunks, providers are asked in pages, large responses are paginated and the export is written in pages. Each of those seams exists because an earlier version fell over on a request limit, a row limit or memory.

Two ways to classify

An LLM reads each search, or embeddings pair it with the closest category by cosine similarity. Each run records which method, the tokens used and what it cost.

Stale results say so

When the inputs change after a job has run, the results it produced are flagged as outdated rather than silently kept, and the job can be re-run on its own.

Honest time ranges

Each provider has its own history limit, 48, 24 or 12 months, and the tool only offers what the data can support.

Batch sizes and timings here are invented. The chunking, paging, watchdog and outdated flags are the real mechanics.

Matches more than names

The filter looks at each row's label, description and metadata, case-insensitively and on any part of the text.

State per view

Each tree view keeps its own filter for the session, so moving between containers and images doesn't lose your place.

Shipped properly

Six commands with menu entries and translations, plus telemetry on filter use, so the maintainers can see whether it earns its place.