agentproto settings

agentproto settings export [--out <file>] [--json]
                            [--include-secrets <profile-id>...] [--passphrase-env <VAR>]
agentproto settings import <file> [--dry-run] [--yes] [--json]
                            [--unseal-passphrase-env <VAR>]

"Bring my main setup" — snapshot this machine's agentproto configuration into one versioned JSON bundle, then apply it to another machine. Both sides are pure local file I/O; nothing round-trips through a daemon or the network.

The bundle covers: installed adapters (slug + version), harness→profile presets, auth-profile metadata (id, endpoint, method — never a credential), named LLM endpoints (~/.agentproto/llm-endpoints.json), imported-MCP pointers (~/.agentproto/imported-mcps.json), and a sanitized slice of ~/.agentproto/config.json. Cron jobs are never included — a kind: "command" cron action can embed an arbitrary shell command, secrets and all, and they live in the daemon's in-memory scheduler rather than a plain file; use cron.md directly on the target machine.

Secret disciplineCopy Anchor Link

Nothing sensitive rides in a bundle unless explicitly asked for:

• Auth-profile credentials are never included by default. Every profile exports as {id, endpoint, method, label?, origin?} — the store pointer (credentialRef) is dropped; it's local-only and meaningless on another machine anyway.
• --include-secrets <profile-id> (repeatable) seals that profile's stored credential into the bundle under a passphrase (--passphrase-env <VAR> — never pass the passphrase as a bare argument; it belongs in an env var, not your shell history). Sealing is scrypt → AES-256-GCM, keyed by the passphrase alone. Sealing to a paired device's public key instead of a passphrase is planned once the device registry exists, but isn't wired up yet.
• An imported MCP server's env/headers can carry literal secrets (an Authorization header, an API-key env var) — only the key names ride in the bundle, mirroring how llm-endpoints.json's apiKeyEnv is already a name, not a value. A stdio MCP's command/args are dropped entirely (not just redacted) when either embeds an absolute local path, since that leaks the source machine's username and directory layout.
• config.json is walked recursively. A key the runtime's own schema marks secret: true is dropped (including a per-profile override like profiles.<name>.tunnel.token, which mirrors the top-level shape one level down), as is any string value shaped like a credential by its key name (token, apiKey, password, …) even where the schema doesn't yet cover it, and any string that looks like an absolute filesystem path or a loopback address (machine-specific, not secret, but still not portable). Every drop is reported by dotted path — never by value.

exportCopy Anchor Link

agentproto settings export
agentproto settings export --out my-setup.json
agentproto settings export --include-secrets work-openrouter --passphrase-env SETTINGS_PASSPHRASE

Wrote settings bundle → ./agentproto-settings-2026-09-27.json

  14 adapter(s)
  8 harness preset(s)
  21 auth profile(s) (metadata only)
  1 LLM endpoint(s)
  14 imported MCP server(s) (env/header VALUES stripped)
  10 config.json key(s) (14 skipped — secret or machine-specific)

A profile id that can't be sealed (unknown id, a source-backed profile with no stored secret, a keychain read failure) doesn't abort the export — it's reported in warnings and simply left out of the bundle.

importCopy Anchor Link

agentproto settings import my-setup.json --dry-run
agentproto settings import my-setup.json --yes
agentproto settings import my-setup.json --yes --unseal-passphrase-env SETTINGS_PASSPHRASE

Plan for this bundle (from other-machine.local, 2026-09-27T16:58:21.652Z):

  auth profiles: 3 to add, 18 skipped
  harness presets: 1 to add, 7 skipped
  LLM endpoints: 0 to add, 1 skipped
  imported MCP servers: 2 to add, 12 skipped
  config.json keys: 4 to add, 26 skipped

(dry run — nothing was applied)

Import is additive only: an entry that already exists locally (same id, or an already-set config.json key) is left completely untouched and reported as skipped, never overwritten. A bundled auth profile with no matching sealed secret — or a bundle with sealed secrets but no --unseal-passphrase-env — is created disabled: a shape-only placeholder an operator fills in afterward with a real credential (agentproto auth login, or the auth_profile_create MCP tool). A harness preset that references a still-disabled profile fails its own validation and is reported skipped with the reason, rather than aborting the rest of the import.

Adapters the bundle names that aren't installed locally are listed with an agentproto install <slug> hint — never auto-installed, since an adapter install shells out and can fetch/run arbitrary code. Cron jobs are never part of a bundle, so import never creates one.

See alsoCopy Anchor Link

• config.md — hand-editing ~/.agentproto/config.json directly; settings import only ever adds a key that's unset there.
• llm.md — the named LLM endpoints a bundle carries.
• install-mcp.md — the imported-MCP store a bundle's mcpServers merge into.
• auth.md — creating/enabling a real credential for a disabled profile stub after import.