ARK · Developers
The layer underneath the products.
Protocols, open tools, architecture notes and the runtime research. Published with its real maturity attached, because an experiment presented as infrastructure wastes someone's week.
Areas
Four surfaces, four different levels of done.
Runtime
ResearchWhere a unit of AI work should execute, decided from capability, privacy, policy, latency and cost rather than from habit. Local, edge and cloud are a routing decision, not an ideology.
Read the thinkingProtocols
Open sourceAgent-facing interfaces: MCP, agent-to-agent patterns, and the machine-to-machine conventions forming around them. We test against the published specs and publish what breaks.
Check a serverOpen tools
Open sourceScanners, catalogues and reference implementations from ARK Labs. MIT licensed, no account, most of them store nothing.
Browse the toolsDocumentation
Open sourceArchitecture notes, schemas and implementation guides, published alongside the code they describe.
View GitHubProtocols
Agent-facing interfaces, tested against the published spec.
The Model Context Protocol changed how servers start up in its 2026-07-28 revision. Existing servers keep working for about a year, which means a lot of them have an expiry date nobody has scheduled.
What a conformance run checks
- Transport
- Handshake
- Capabilities
- Schemas
- Tools
- Report
“It works in one client” stopped being evidence the moment enterprise platforms started publishing tool surfaces by the dozen. The scanner is open source and runs from the CLI, so it can sit in CI rather than in someone’s memory.
Runtime
Where should a unit of AI work execute?
Local, edge and cloud are usually argued as an ideology. They are a routing decision, and the inputs to it are boring and measurable.
- Capability
- Does the work actually need a frontier model, or is a smaller local one sufficient and more predictable?
- Privacy
- Can this data leave the boundary at all? Some of it cannot, and that decides the answer on its own.
- Policy
- Which jurisdiction, retention rule or disclosure duty applies to this unit of work?
- Latency
- Is a round trip acceptable, or does the work sit inside something a person is waiting on?
- Cost
- What does this cost per verified outcome, not per token, once retries are counted?
We are publishing the thinking, not selling a runtime. When there is something to run, this page will say so.
Open tools
MIT licensed, no account, most store nothing.
6 of the ARK Labs projects ship their source. Clone them, read the rules, disagree with the rules, open a pull request.
AI Workflow Templates
Open sourceReady-to-run AI workflows you can copy, wire up, and ship today.
Conventions
The agent architecture we build against.
Scope, identity, access, action, stop. Five things an agent needs settled before it is allowed to touch a production system.
- Scope
- Identity
- Access
- Action
- Stop
The full framework — goal, trigger, state, boundary, evidence and owner, with worked examples of a bad and a good version of each — is published free.
Building the agent side of a real system?
If you are implementing protocols, tool authority or verification against production systems, we would rather talk to you than market at you.