๐Ÿ”Œ Custodian Integrations ยท One kernel, every agent

The kernel doesn't care
which agent you run.

Custodian sits outside the model, not inside its prompt โ€” so the same policy, the same fail-closed adapters, and the same hash-chained receipts cover any agent you point it at. Codex Guard and Talaria (for Hermes) are live today. Claude, Gemini, and opencode integrations follow the same pattern.

Live today

Two agents, governed now

Each integration is a thin, agent-specific adapter over the same kernel primitives โ€” bands, caps, guard adapters, kill switch, hash-chained receipts. Nothing agent-specific lives inside the kernel itself.

Codex Guard
for OpenAI Codex coding agents
โ— Live

A capability firewall for coding agents. Codex can inspect, test, and edit freely inside an approved workspace โ€” but credential use, network calls, destructive commands, production changes, money movement, and governance edits stop at a human-approval boundary the model cannot grant itself.

  • Repo-local plugin + dependency-free MCP server (guard_action, verify_receipts)
  • Doesn't trust the model's own risk label โ€” known tool/command shapes independently raise the minimum risk class
  • Action-bound, expiring, single-use human approvals
  • Value-free HMAC hash-chained decision receipts
Talaria
for Hermes Agent ยท named for Hermes' winged sandals
โ— Live

The integration surface between a local Hermes Agent and everything the Custodian kernel enforces โ€” secrets, files, tools, prompt-injection, and spend anomalies, all checked mechanically below the model, not asked of it. One command installs a kill switch and a tamper-evident audit trail.

  • custodian setup --profile hermes โ€” installs and wires every hook
  • Denies reads/writes to ~/.ssh, ~/.aws, ~/.gnupg, *.env, *.pem
  • Caught 5/5 live secret-leak evasion attempts in testing
  • Guard-level checks out of the box; deep skill integration adds kernel-enforced spend caps + credential brokering
Coming next

Same kernel, more agents

Each new integration is the same shape: a thin adapter that translates one agent's tool-call surface into the kernel's action schema. No agent gets special-cased inside the kernel.

Claude
Planned
Gemini
Planned
opencode
Planned
Your agent
Same adapter pattern โ€” ask
One policy

Governance doesn't fork per agent

Bands, spend caps, the kill switch, and the guard adapter catalog are defined once. Swap Codex for Claude or Hermes for Gemini and the policy stays put โ€” only the thin adapter translating that agent's tool calls into kernel actions changes.

One audit trail โ€” in progress

One ledger, one agent at a time today

Codex Guard's decisions are hash-chained, tamper-evident, and readable live from the Live Console today. Talaria keeps its own separate audit trail for now โ€” unifying both into the one ledger every agent writes to is active, not-yet-shipped work, not a retrofit we're claiming already happened.

Bring your own agent.

The kernel was built to sit outside any single model or vendor. If you're integrating an agent that isn't listed above, the adapter pattern is documented and the kernel doesn't need to change.