Ready and not ready
LOOPmKit is early. This page says what was checked, what is in progress and where the limits are. We do not claim that an AI company runs itself with it.
83Implemented and tested
Ready
The rest of the feature list is implemented with automated tests. That is a statement about the code, not a promise of fit for your case. Features
Checked on a running stack
These were run against a real API and database, some with a scripted model and no outside actions.
- Launch parametersBuilder and triggersChecked on a running stack
- SkillsBuilder and triggersChecked on a running stack
- Ask a person mid-runApprovals and policyChecked on a running stack
- Product tasksAPI, webhooks and MCPChecked on a running stack
- Works at phone widthWeb interfaceChecked on a running stack
- Company wizardTemplates and company setupsChecked on a running stack
- Product passportTemplates and company setupsChecked on a running stack
17In progress
In progress
Partly built, or finished in code but not switched on yet.
- Circuit breaker per providerModels and routingIn progress
- Token counting before sendingModels and routingIn progress
- Embedded price listCosts and budgetsIn progress
- Subscription providersCosts and budgetsIn progress
- "Memory required" policyMemoryIn progress
- Per-operation risk contractsConnectors and toolsIn progress
- Social connectorsConnectors and toolsIn progress
- Logged-in browser profilesBrowser, desktop and code executionIn progress
- Content guardrailsApprovals and policyIn progress
- Tool results as untrusted dataApprovals and policyIn progress
- Prompt optimisationEvals and qualityIn progress
- Error feedEvals and qualityIn progress
- OpenTelemetry tracingObservabilityIn progress
- LOOPmKit as an MCP serverAPI, webhooks and MCPIn progress
- Voice commandsAPI, webhooks and MCPIn progress
- Phone agentAPI, webhooks and MCPIn progress
- Owner cockpitWeb interfaceIn progress
On the roadmap
- Full development cycleTemplates and company setupsOn the roadmap
- Kubernetes and HelmDeploymentOn the roadmap
Every feature as one square, colored by status. Implemented and tested: 83 / 102.
- Checked on a running stackChecked on a running stack with real services (some with a scripted model).
- Implemented, with automated tests. No mark.
- In progressPartly built, or built but not switched on yet.
- On the roadmapNot built yet.
Known limits
Stated plainly, as in the product's own notes.
- One operator. There are no multi-tenant organisations and no roles and permissions between users.
- The goal of an AI company is not reached. Six of the ten acceptance criteria are still open: employee management with a test run, connection authorisation checks, the full development process, support, the owner cockpit, and full test plus live acceptance.
- A configured employee means roles and permissions exist. Repository access, execution tools, outside accounts and day-to-day acceptance are not ready.
- Employee inbox agents answer from the model alone and do not execute code.
- No run has been done on real Claude Code or Codex subscriptions yet. The code execution path was tested with test doubles and disposable containers.
- Desktop control works only on macOS hosts, with the bridge set up by hand.
- Live acceptance of the mail and chat templates is partial. The Gmail connector the template relies on is archived upstream.
- Approving each dynamic mail action separately is not supported by the runner yet.
- Cost figures depend on the model prices configured on a provider. An unknown price counts as zero in the metric, and tiered, batch and regional prices are not modelled.
- The document template does no upload or OCR of scanned PDFs. The page check template is not a change monitor.
- Social connectors: live acceptance and automatic sign-in are open.
- In the container presets Chromium runs without its own sandbox, because the default Docker profile blocks it. A fix is described but is not the default.
- Several finished features are not switched on in the author's own setup yet: the MCP server, tracing, content guardrails, risk-contract backfill, the price list, the error feed and mandatory memory.
- Not planned for now: a semantic cache and hedged requests.
- The repository is private and the license is undecided. Terms are not set, and there is nothing to download.
Early access
If you want to try it, the way in is a call.
Book a 20-minute call