Start small. Make it repeatable.

Run the platform

These instructions describe a deployment from an existing copy of the application source. If you already have a running instance, start by connecting a model.

From a checkout, Docker and Make build and run the plane. Go and Node are provided inside the build containers. Linux is the supported host for the execution sandbox.

export ALTHERIUM_TOKEN=$(openssl rand -hex 16)
make up
make logs

Open localhost:8080. Local data lives in ./.altherium. Keep the generated token to access the instance. In token mode, an empty token leaves local access open.

SQLite is the default. For PostgreSQL with pgvector, use make up-pg; stop that deployment with make down-pg. Files and secrets still need the plane data directory.

Connect a model

Choose a supported cloud provider in the model and connection settings, or add a local inference node. A GPU cluster is not required to use cloud models.

For local inference, open Nodes → Add node and use the generated enrollment instructions. The one-use bootstrap token enrolls the node; the plane distributes the inference runtime. Node-to-plane traffic uses mutual TLS.

Model availability and context limits depend on the provider and the instance configuration. Review the selected model rather than assuming every provider has the same capabilities.

Give the task context

Create a project with the appropriate personal or team scope. Add files or attach a Git repository, then open a conversation connected to that project.

Connect only the tools the task needs. A skill supplies a reusable procedure; knowledge supplies searchable documents; memory holds scoped facts. They serve different purposes.

Use people’s connections for interactive work and configure the team service account for scheduled work. Service accounts have their own identity and permissions, without global administration.

Turn a task into a workflow

Start from one task with a result you can inspect. Define its agent steps in the workflow editor or import a YAML definition. Add dependencies to express the intended order.

Test it manually before adding a schedule. Inspect the run and the output of each step. Durable progress lets workflows resume after a pause or restart; it does not make external actions automatically idempotent.

Decide where autonomy stops

Attach the relevant policies and checkpoints to the workflow. Define what permits progress, what needs a human decision and what must be blocked. Use verified signals where the policy requires evidence.

Review behavior in shadow mode, inspect recorded gate decisions and compare a draft policy against the available history. Approvals, spend limits and brakes complement those workflow controls.

The repository documentation is the detailed reference for deployment and operation. Start with a supervised workflow and expand its autonomy based on the results.