Discovered from GitHub
Repo scanning proposes candidates with AI-drafted descriptions and owners inferred from commit history — no catalog-info.yaml to commit to every repo. Nothing enters the catalog silently.
Your service catalog, without the modelling project.
Connect GitHub. Services, owners, tiers and 32 scorecard rules are already modelled — first scorecards in minutes, and agents that can act on the catalog safely.
Self-serve signup. Free plan: no card, no LLM key.
Works with your stack
Onboarding
Your catalog builds itself
What you actually run becomes the catalog: proposed for you, confirmed by you. Connect GitHub and Kubernetes — first scorecards in minutes, not weeks.
Discovered from Kubernetes
Team planOne Helm chart discovers every workload: outbound-only, read-only, no per-service setup.
Drift caught in both directions
Running workloads missing from the catalog get flagged, and so do catalog entries with nothing behind them.
No schema design project
Services, databases, containers, ingresses and more are already modelled.
Also included
- Dependency graph
- Scaffolding templates
- Self-service actions
- DORA metrics
- Cloud cost attribution
- Per-agent LLM spend
Compliance
Compliance signals from day one
Security, Ownership and Delivery start scoring the moment you connect GitHub. No rules to write first.
- 32Built-in rules
- 4Rule categories
- 3Tiers: bronze, silver, gold
Weighted by criticality
Your payment service is held to a higher bar than an internal tool.
The next fix, named
Every scorecard points at the single rule standing between you and the next tier.
Tier changes on the pull request
They post as GitHub checks, so a drop is caught before it merges.
How it works
Take a tour
AI Readiness
Ready for an agent
Before you point an agent at a service, know whether it's ready for one.
Scored like everything else
Rules that check whether a service is ready for an agent to work on, so you know where you can safely point one.
- Agent instructions present
- Bootstrap commands listed
- Generated code marked
- Flaky tests under control
Agents
Point it at real work
Purpose-built agents — or build your own from scratch.
Service Explorer
Answers questions about the catalog — ownership, compliance posture, repos — so nobody's hunting through Slack for who owns what.
Task Manager
Finds and runs the right self-service action for a day-2 task — scaling, restarts, provisioning.
Incident Manager
Summarises a newly opened incident with the service and owner already resolved, the moment it opens.
Compliance Remediator
Investigates a newly failing rule and proposes the exact pull request that fixes it.
Governance
Agents reach only what you allow
Decisions route to a human when they should, and every step stays on the record.
Per-tool access levels
Set on every tool of every connected MCP server.
- Always allowRoutine, low-risk tools
- Ask once per conversationConfirm, then continue
- Ask every timeFor the irreversible ones
- BlockThe agent never sees it
An allowlist you control
Agents run with only the tools you grant. Nothing else exists to them.
Approval that can't be self-served
Any action can require human approval, and the approver can never be the person who requested it.
Approvals expire, policy re-checks
Block a tool while a request is waiting and that request can no longer be approved. Policy applies at decision time, not submission time.
Append-only audit log
Every agent action, approval and tool change is recorded, enforced at the database. Nothing gets removed, only appended to.
MCP security
Tools locked to what you approved
A tool's description is executable code.
Real incidents have already exploited exactly this — credentials silently exfiltrated, malicious changes auto-executed on the next interaction, after the tool had already been reviewed and approved. A change after sign-off doesn't require a new approval. Unless you pin it. (CVE-2025-54135, CVE-2025-54136)
- Reworded description, vetted serverNotifies
- Any change at all, unvetted serverSuspends the tool
- New parameter, widened bound, opened fieldSuspends the tool
Pinned at review
The definition a human approved is the only one your agents are ever shown.
Drift detection down to the schema
Checked automatically about every 10 minutes, or on demand. We diff the whole input schema — types, required fields, enums, bounds — not just the description.
Suspended means suspended
Unattended automation is refused outright; chat needs per-call approval — until someone reviews the diff and re-pins.
A review queue that ranks itself
What actually needs a human rises to the top, identical changes group together, and every diff sits side by side.
Withhold from every agent
One click holds a tool back from every agent on the connection, without waiting for the next drift check.
Pricing
Priced for builders. Everyone else is free.
You pay for builders. Everyone who reviews, approves or audits is free on every plan — guardrails shouldn't cost extra.
Free
$0
Evaluate with your whole team. No card, no LLM key.
- 1 builder, plus unlimited free viewers and approvers
- 15 agent runs/mo
- 20 services in the catalog (unlimited discovered candidates)
- 2 agents
- 1 MCP tool server connection
- All 32 rules, all categories, bronze/silver/gold scoring
- DORA metrics and dashboards
- Assistant chat, included.
- 30-day audit retention
- Role-based access control
Team
Featured$29/builder/mo · 5 builder minimum
For teams running agents in earnest.
- 20 agent runs/builder/mo · unlimited with your own LLM key
- 50 services
- 5 agents
- 5 MCP tool server connections
- 1 Kubernetes cluster
- Customisable role permissions
- Pull-request tier gates
- Unlimited custom rules
- 90-day audit retention
Custom
Let's talk
Limits, audit retention and onboarding shaped around you.
- What counts as a builder?
- Anyone who changes something — the catalog, the rules, the agents. Reading, reviewing and approving are free and unlimited.
- Every organisation starts on a 14-day trial of Team, with 5 builders and 50 agent runs. No card required. You drop to Free automatically when it ends, so there is nothing to cancel.
- An administrator assigns the approver role, so the person who releases what an agent does is never the person who built it.
Connect GitHub. See your catalog in minutes.
Services, owners, dependencies and scorecards, proposed from what you actually run and confirmed by you. Then point an agent at the services that are ready for one.