Build on Agentware · Haikei docs

Use the Agentware SDK to add Kei governance to a custom agent harness. Agentware provides the middleware, audit, guardrails, and Kei policy integration; Kei is the metadata-only control plane that holds registration, policy, and audit metadata. Provider payloads, customer content, and credentials never enter Kei.

For a first look at what agentware does, see Govern an agent with agentware.

Six harness types#

Kei has six user-facing harness types. The SDK (Agentware) type covers harnesses built on the Agentware SDK:

Harness type	Integration path
SDK (Agentware)	`KeiProxyEvaluator` → `kei-proxy authorize`, fail closed. Covers pydantic-ai, Gemini SDK, AWS AgentCore, and bespoke loops. API kind `custom`.
Claude Code	Native permission files + hook.
Codex	Native permission files + hook.
OpenCode	Native permission files + hook.
Pi	Own package + local-socket adapter.
Open WebUI	Own Python library (`openwebui-kei`), not the SDK path.

Six-step path#

Bring up an Agentware SDK harness in this order:

Install the Kei CLI and sign in — curl -fsSL https://haikeilabs.com/install.sh | sh, then kei login. You need an organization owner or admin to approve the device authorization flow.
Create a runtime installation — kei bot init creates pending installation metadata and provisions a runtime credential. The credential is shown once; deliver it directly into your approved secret manager.
Register the harness — kei harness add --kind custom --installation INSTALLATION_ID registers the SDK harness. The CLI resolves the assigned agent for that installation.
Register tools — Use ToolSpec with a harness_executor route over the runtime socket (POST /v1/tools:register), or maintain a tools.json manifest and sync it with kei harness sync --harness custom --manifest tools.json. The manifest is tool metadata, not a permission grant.
Start the runtime daemon — kei-proxy serve runs the policy enforcement point on a local Unix socket. The SDK discovers it through KEI_RUNTIME_SOCKET_PATH.
Run the harness — Each governed tool call goes through kei-proxy authorize. An explicit allow lets the tool run; a deny returns reason_code and hint; a socket or protocol failure fails closed.

Compatible versions#

agentware SDK	kei-proxy (minimum)	kei-cli (minimum)
0.11.0	v0.3.0	0.2.0

Language guides#

Next steps#

Runtime setup — connect a harness to a local kei-proxy daemon.
Policy and audit — create policies and view audit entries.
Deny reason codes — reference for every reason_code and its hint.

For the complete API reference, see the Agentware SDK repository.