A connection is a saved, reusable link to an external service, with its credentials stored securely. Create it once, reference it from any agent or workflow.
Providers — the brain your agents think with: Anthropic, OpenAI, LM Studio.
Integrations — what your AI acts on: Postgres, or any OpenAPI service as typed actions.
Gateways — an optional proxy (Apigee) for monitoring and policy.
Two first-class ways to build, and they’re peers. Use either, or combine them — an agent can even be a tool inside a workflow.
Agents — instructions, a model, and the tools they’re allowed to use.
Workflows — a visual graph of nodes, wired output to input on the canvas.
Tools — an integration action, a whole workflow, or another agent.
Deploying freezes the current draft into an immutable, numbered version. A deployment is a movable pointer to one version — promote forward, or roll back anytime.
Versions — immutable and numbered, so every run traces to an exact definition.
Deployments — promote a version live; only deployed work can be run.
Rollback — point back to an earlier version instantly and atomically.
A trigger is what starts a run. Talk to it, put it on a timetable, or call it from your own systems — webhooks and events are on the way.
Compass Chat — a conversational interface; attach files, watch responses stream, traces one click away.
Schedule — a cron-style timetable for recurring work.
API — run from anywhere in your stack, over the REST API or SDK.
Letting AI act is a decision, not a default. Compass puts the controls where they belong — approval gates on the actions that matter, hard limits on spend, and roles that decide who can do what.
Approvals — gate any tool or action behind human sign-off; the execution pauses until someone decides, then resumes exactly where it stopped. Every decision is recorded — who, what, when.
Spend limits — budgets on an organization, project, or connection — hit one and new work is blocked before the model is even called.
Roles — admins, developers, and read-only viewers decide who builds, who runs, and who only watches.
Every agent and workflow runs on a durable execution engine — the layer most AI platforms don’t have. Each execution is checkpointed and orchestrated step by step, so it survives failures, scales to heavy workloads, and never does the same work twice.
Durable — a failed step or a restarted machine resumes from its last checkpoint, not from scratch.
Scalable — thousands of concurrent executions and long-running jobs — point an agent at a list and it fans out, one execution per item, at a concurrency you set.
Isolated — each execution runs against its exact deployed version, on its own, with automatic retries on transient failures.
Every execution records a trace of every step it took. Watch it stream live — node by node, token by token — or explore the whole tree afterward to understand and debug.
Executions — input, output, tokens, duration, trigger, and which version executed.
Traces — every node and tool call, nested, with inputs and outputs.
Debug — replay node by node to find exactly where something broke.