Connect each account separately and give it a recognizable name, such as Gmail Research, Slack Engineering, or Anthropic Coding. Each connection has its own permanent ID, owner, status, and tool permissions.
Connect and select an account
- Open Connections → Add connection and choose the integration.
- Name the account and complete its authorization. Repeat for another account of the same integration.
- Approve tools under Tools, then add rules under Access. Presets and runs can inherit eligible connections, select specific tools, or explicitly choose none. See connector access rules.
Edit changes the display name without changing its ID. The account identifier shown on an authorized app connection is Composio's verified connected-account ID; it is not an inferred email address. Connecting another account never changes an existing agent's selection. Reconnect targets the same provider account. To switch to another account, create another connection and explicitly select it.
Connections belong to their authorizing user. Tool access can target the organization, a workspace, a preset, or an exact workspace + preset pair; model keys and Claude subscription configurations remain owner-bound. Disconnect removes local authority immediately. Upstream revocation can require a separate provider action; see connection security.
Custom model hosting connections
Hugging Face, RunPod and external model connections are owner-bound hosting credentials. Open Connections → Add connection → Model hosting, choose the connection type, enter its account API token or endpoint settings, then Save connection. The same form is available during Model deployments setup. Saving credentials starts no compute and performs no inference. Protected configuration files with the connection CLI use the same API.
For RunPod model hosting, use RunPod account with your RunPod API key. No app OAuth or connector enablement is required. Searching Apps for RunPod opens this hosting path; Connect RunPod app tools instead deliberately selects the separate tool integration, which requires operator enablement. An existing RunPod app-tools card offers Use for model hosting to create the correct credential without changing its original identity or grants.
Connections → Edit accepts a masked replacement token; macrofold connection update ID --config-file private.json uses the same API. A RunPod replacement key must still reach every existing exact endpoint; unresolved creation blocks rotation. Disconnect closes ordinary use while retaining narrow cleanup authority for owned infrastructure. Hosting connections have no MCP tool permissions. See custom model setup and operation.
Save authentication in a preset
An agent preset saves its harness, model, funding mode, provider connection, instructions, and optional tool-selection defaults. The dashboard preset editor lets you choose a named API-key connection. API and SDK clients use provider_connection_id for model authentication and connection_grants for app tools.
For example, using an existing Anthropic API-key connection:
from macrofold import Macrofold
macrofold = Macrofold()
preset = macrofold.agents.create(
name="Coding",
harness="claude-code",
model="YOUR_ENABLED_ANTHROPIC_MODEL",
billing_mode="byok",
provider_connection_id="YOUR_ANTHROPIC_CONNECTION_ID",
)
run = macrofold.runs.create(
workspace_id="YOUR_WORKSPACE_ID",
agent_id=preset.id,
prompt="Update the workspace notes.",
)
for text in macrofold.runs.stream_text(run.run_id):
print(text, end="", flush=True)
macrofold.close()
From a linked terminal directory, use macrofold run "Update the workspace notes." --agent AGENT_ID. The preset supplies authentication and limits. Explicit timeout/budget flags override only the corresponding limit. Use a new session when changing authentication; continuing an existing session keeps its original configuration.
TypeScript, Python, Go, Rust, and Java expose the same generated connection and preset fields. See SDKs and the API reference.
Local simulation can select an owned API-key connection to exercise this journey without reading or sending its key. Native execution requires a model from that key's provider; a simulation preset must select a compatible enabled model before switching to native execution.
Claude subscription availability
You can save multiple named Claude subscription configurations, such as Research, Coding, and Personal, and reference them from Claude Code presets using billing_mode="subscription". These configurations are not authenticated accounts: subscription login and execution are currently unavailable. They stay pending and do not collect Claude login tokens or credential files. Anthropic API-key execution remains available through the existing Claude Code harness.
This gate applies to every invocation surface, including simulation. A rejected subscription run does not start a sandbox, reserve funds, or silently use an API key. The maintainer acceptance record distinguishes implemented configuration from the remaining authentication/runtime work.
Optional backup policy
In a Claude configuration's Edit dialog, you can save an owned Anthropic API-key connection and a positive per-run API spending limit. The default is disabled. This policy remains inactive while subscription execution is unavailable.
The API field is api_fallback: set enabled to true with connection_id and max_cost_micro_usd, or set enabled to false to clear the selection. One million micro-USD equals $1. Omitting the field during an edit preserves its current setting.
The policy permits paid fallback only for confirmed quota exhaustion before native execution starts. Unknown quota, ordinary authentication errors, temporary failures, and ambiguous launches do not authorize spending. After execution starts, quota exhaustion requires an explicit continuation over preserved files/history; the original prompt must not be restarted automatically. Infrastructure charges are separate from any API fallback allowance.