How Kei governs coding agents

Kei supports six harness types: Claude Code, Codex, OpenCode, Pi, Open WebUI, and SDK (Agentware). Claude Code, Codex, and OpenCode are desktop harnesses; Pi uses its own package and local-socket adapter; Open WebUI uses its own integration; other agents integrate through the Agentware SDK. Each developer machine has one runtime installation, and each harness session is associated with it. Kei renders eligible command policies into desktop editor settings; each editor makes its own native decision (allow, deny, or ask), and Kei audits the session.

How policies are created and distributed#

mermaid

flowchart TD
  subgraph Admin["Org admin / owner"]
    A1["Console → Policies<br/>(Create agent command policy)"]
    A2["kei policies create<br/>(kei login)"]
  end
  subgraph CP["Kei control plane"]
    G["Groups & Roles<br/>admins ⊃ developer …"]
    P["Policy<br/>src: group:developer · user:x · harness:claude_code · *<br/>dst: shell:git · skill:wiki · path:/repo<br/>effect: allow|deny · priority"]
    R["Policy revision"]
    B["Policy bundle per installation<br/>version · digest · expiry"]
  end
  A1 --> P
  A2 --> P
  G -->|"effective group and user grants"| B
  P --> R --> B
  subgraph Mac["Developer machine: one runtime installation"]
    S["kei runtime service<br/>(heartbeat, policy refresh, audit upload)"]
    ST[("cached policy bundle")]
    SYNC["kei harness sync"]
  end
  B -- "fetch with the runtime credential" --> S --> ST --> SYNC
  SYNC -- "allows / denies / Skill()" --> N1["Claude Code settings.json"]
  SYNC --> N2["Codex rules (allow / forbidden)"]
  SYNC --> N3["OpenCode permission block"]
  SYNC -. "automatic re-sync on bundle change (planned)" .-> SYNC

Admins create policy in the console or with the Kei CLI, and Kei records each change as a policy revision.
Kei distributes a versioned, expiring policy bundle to the machine’s single runtime installation.
kei harness sync translates the bundle into native permissions for the installed harnesses.
The runtime evaluates group, user, and autogroup:member policy sources for governed calls. For desktop command policies, a group: source renders when it matches one of the bundle subject’s groups, and a user: source renders when it matches that subject. A non-matching source is skipped; policies for user and group sources are not rendered for other subjects. autogroup:member resolves at runtime for every organization member and is never stored. Automatic re-sync is planned, as shown by the dotted arrow.

Bring your own editor#

mermaid

flowchart TD
  U["Developer, signed in as themself<br/>member of group:developer"] --> I["One runtime installation per machine<br/>(one Agents card)"]
  subgraph Harnesses["Six harness types"]
    C["Claude Code"]
    X["Codex"]
    O["OpenCode"]
    K["SDK (Agentware)"]
    P["Pi"]
    W["Open WebUI"]
  end
  I --> C & X & O & K & P & W
  C -- "native permissions + report-only Kei hook" --> D1{"Native decision<br/>allow / deny / ask"}
  X --> D1
  O --> D1
  K -- "KeiProxyEvaluator → kei-proxy authorize (fail closed)" --> D2{"kei-proxy decides"}
  P --> P1["Pi-specific local-socket adapter"]
  W --> W1["Open WebUI's own image and tool packs"]
  D1 -- "report-only event: kind + session" --> AU["Audit upload"]
  D2 --> AU
  AU --> UI1["Audit Logs (harness kind, session)"]
  AU --> UI2["Agents card: live sessions"]
  M["Model choice per harness"] -. "switching models through Kei (future)" .-> C

One installation supports multiple harnesses and sessions on the same developer machine.
Supported editors apply eligible Kei-generated native permission settings and make their own native decision (allow, deny, or ask). The audit hook reports activity and does not decide or block calls.
SDK (Agentware) integrations use KeiProxyEvaluator to call kei-proxy authorize, which fails closed when it cannot authorize a request. Start with Build on the agentware SDK.
Pi and Open WebUI remain separate harness types with their own integrations; see Pi setup and Open WebUI with Kei.
Kei receives audit events with the harness kind and session, and shows activity in Audit Logs and the Agents card.

What’s planned#

Kei plans to automatically re-sync harness permissions when a policy bundle changes and to support switching models per harness through Kei. These capabilities are planned and do not have promised release dates.

How Kei governs coding agents

How policies are created and distributed#

Bring your own editor#

What’s planned#

Learn more#