mngr_imbue_cloud

Provider backend plugin and CLI for Imbue Cloud (the imbue-team-hosted leasing service that talks to remote_service_connector).

The plugin owns these concern areas, all reachable only through mngr commands:

• auth — SuperTokens session: signup, signin, oauth, refresh, signout, status
• hosts — lease/release/list pre-provisioned pool hosts
• keys — LiteLLM virtual key management (mngr imbue_cloud keys litellm ...)
• buckets — R2 bucket + scoped-key management (mngr imbue_cloud bucket ...)
• tunnels — Cloudflare tunnel + service + auth-policy management
• admin pool — operator-only pool provisioning (Vultr + Neon)

Configuration

Each signed-in account is its own provider instance entry in ~/.mngr/config.toml:

[providers.imbue_cloud_alice]
backend = "imbue_cloud"
account = "alice@imbue.com"
# connector_url is optional; defaults to the prod URL.

The default connector URL can be overridden via the MNGR__PROVIDERS__IMBUE_CLOUD__CONNECTOR_URL environment variable.

Sign in

mngr imbue_cloud auth signin --account alice@imbue.com
# or browser-based OAuth:
mngr imbue_cloud auth oauth google --account alice@imbue.com

Token state lives at <default_host_dir>/providers/imbue_cloud/sessions/<user_id>.json.

Create an agent on a leased host

Use the standard mngr create pipeline -- the plugin's provider backend runs the lease + SSH bootstrap inside create_host, and the rest of mngr's create flow adopts the pool's pre-baked agent under your chosen name:

mngr create my-agent@my-host.imbue_cloud_alice --new-host \
    -b repo_url=https://github.com/imbue-ai/forever-claude-template \
    -b repo_branch_or_tag=v1.2.3

--build-arg KEY=VALUE flags become LeaseAttributes (repo_url, repo_branch_or_tag, cpus, memory_gb, gpu_count); the connector matches them via JSONB containment against the pool host's attributes row. Any other -b entry (e.g. --file=Dockerfile, .) is forwarded verbatim as a build arg to the slow-path container rebuild.

Fast path vs. slow path (fast_mode)

mngr create against imbue_cloud can land on a pool host two ways, selected by -b fast_mode=<require|prevent>:

• fast_mode=require -- the fast path. Lease a pool host whose attributes row exactly matches and adopt its pre-baked system-services agent. The pool host is fully pre-provisioned, so mngr's create pipeline only writes the agent env file (and patches the claude config when an ANTHROPIC_API_KEY lands in env) before starting the tmux session. If no exact match is available, this raises FastPathUnavailableError rather than falling back.

• fast_mode=prevent -- the slow path, and the default. Lease any adequately-sized available host (resource attributes only; repo_branch_or_tag/repo_url are dropped), destroy its baked container, and rebuild it from the FCT Dockerfile via the shared mngr_vps_docker setup path. mngr's standard create pipeline then does full client-side setup -- exactly as if this were a fresh OVH host. The rebuilt container keeps the lease's pre-baked host_id/agent_id so identity stays aligned with the connector's lease row.

The slow path needs a usable build context: run mngr create from (or --project at) a forever-claude-template checkout whose imbue_cloud create template supplies the Dockerfile build args. The logs state which path was taken (imbue_cloud[...] FAST PATH vs SLOW PATH).

The client owns the machine the moment the connector marks it leased: if any step after a successful lease fails, the lease is released back to the pool (no data wipe -- nothing sensitive exists yet) before the error propagates.

minds drives this automatically: it tries fast_mode=require first and, on FastPathUnavailableError, retries with fast_mode=prevent.

When the pool is genuinely empty, even the relaxed slow-path lease returns ImbueCloudLeaseUnavailableError (distinct from FastPathUnavailableError).

Destroy / delete / stop

• mngr destroy <agent> is terminal for imbue_cloud-leased hosts. It (1) stops + removes the workspace container, drops the per-host docker named volume, deletes the per-host btrfs subvolume under /mngr-btrfs/, runs docker system prune -a -f --volumes, and wipes /root + /tmp (keeping /root/.ssh/authorized_keys so the pool-management ssh path keeps working through cleanup); then (2) releases the lease back to the pool via the connector. Privacy-first ordering: the user's data is gone before the connector flips the row to released. The underlying VPS is destroyed later by apps/remote_service_connector/scripts/cleanup_released_hosts.py.
• mngr delete <agent> (or mngr imbue_cloud hosts release <lease-id>) runs the same flow; it's the path mngr's GC takes after the destroyed-host grace period. Safe to re-run on an already-released lease.
• mngr stop <agent> is the "I'll resume this workspace later" path: docker stops the container on the leased VPS, preserves the lease + the on-disk volume, and mngr start <agent> brings the same workspace back up on the same VPS.

Buckets

Create an R2 bucket (for storing files remotely) and mint scoped S3 keys for it. Requires a paid account. Each bucket is isolated (think one per host); the server derives the real R2 name as <user_id_prefix>--<your-name>.

# Create a bucket; emits {bucket, key} where key includes the one-time secret.
mngr imbue_cloud bucket create my-backups --account alice@imbue.com

# List / inspect / destroy (destroy refuses a non-empty bucket).
mngr imbue_cloud bucket list
mngr imbue_cloud bucket info my-backups
mngr imbue_cloud bucket destroy my-backups

# Mint additional keys, scoped read-only or read-write, to hand to agents.
mngr imbue_cloud bucket keys create my-backups --alias agent-ro --access read
mngr imbue_cloud bucket keys list                # all keys across buckets
mngr imbue_cloud bucket keys list my-backups     # just this bucket's keys
mngr imbue_cloud bucket keys destroy <access-key-id>

The emitted credentials (access_key_id, secret_access_key, s3_endpoint, bucket_name) are standard S3-compatible credentials -- point any S3 client at the endpoint. The secret is shown only once at creation and is never stored by the service.

Pool admin

mngr imbue_cloud admin pool create --count 1 \
    --version v1.2.3 \
    --workspace-dir ./forever-claude-template \
    --management-public-key-file ./id_ed25519.pub \
    --database-url "$NEON_DB_DIRECT"