Technical configuration

Models and routing

• Auto-routing (recommended).Set your organization's default to Auto and the router picks the best model per task within the limits you set, escalating off any unavailable provider automatically so runs still complete.
• Per-agent override. Pin a specific model to a specific agent when you want to — your coordinator on one model, a high-volume drafting agent on a cheaper, faster one, in the same team.
• Exclusions. Rule out a whole provider or a single model org-wide; the router respects your exclusions.
• Supported providers. Anthropic, OpenAI, Google, xAI, Mistral, Together AI, Vertex AI Gemini, Azure OpenAI, and OpenRouter — 9 providers spanning 30+ model catalog entries from economy to frontier-class.

Bring your own key (BYOK)

Supply your own API key for any supported provider and that provider's calls dispatch through your account, billed by the provider directly with no platform markup. This is the recommended path when your compliance team holds a Data Processing Agreement with the model provider. BYOK is available for Anthropic, OpenAI, Google, xAI, Mistral, Together AI, and OpenRouter; Vertex AI Gemini and Azure OpenAI are platform-funded routes. You can also rely on platform-funded keys as a fallback: on Pay-as-you-go these meter at the provider's cost plus a 10% margin, and Pro-plan overage bills at model cost with no markup. The in-app picker shows the live per-model rate before you commit.

Channels

Channels are configured per team, with per-agent overrides:

• Free: Dashboard chat and Telegram.
• Pro: Email (your own Gmail or Microsoft 365 inbox via OAuth), Slack, Twilio SMS / WhatsApp / Voice, and generic webhooks. A GroundPound-hosted mailbox on your verified domain is built but coming soon; it is not yet materialized for tenants.

Knowledge and external sources

• Direct upload of documents, with bulk import (review suggested titles and tags, then approve in a batch).
• External connectors — Google Drive, OneDrive, Dropbox, Box — connected via OAuth and kept in sync, with per-folder scoping down to a single agent.
• Web Research — native current-web search is part of the core baseline toolkit for every agent and is separate from optional paid search add-ons.
• Three scopes. Knowledge attaches at the organization, team, or agent level and inherits downward; specific scopes win where they need to.

Tools and skill packs

• MCP servers add tools (the 58-row curated catalog or your own server URL); tool names are namespaced per server to avoid collisions. Credential-gated entries cannot be activated until an org credential or OAuth connection is attached.
• Skill packs add house style and task plans in markdown. The platform currently ships 91 skill packs, stitched into agent instructions in scope order.
• Both attach at org, team, or agent scope.
• OpenClaw import. Already running agents on OpenClaw? A team-scoped import previews the full mapping first, then applies it — nothing lands without your review.

Approvals and policy

• Irreversible, externally-visible actions are approval-gated by default.
• Operators can adjust which categories require approval; high-stakes categories (such as trading orders) cannot be set to unattended scheduled approval.
• The highest-stakes categories are backed by an independent model verification pass. Your policy sets when it fires (for example, trading orders above a notional threshold you choose) and how deep the review goes — from a single second-opinion reviewer to a parallel voting panel of specialists — and the verdict lands on the approval card.
• Outreach approvals pass an integrity gate: every recipient must carry provenance (where the contact actually came from), claimed touch counts are verified against the real send ledger, and a request that re-files after you dismissed it with a note is blocked.
• See the Operator playbook for recommended approval policy as trust accrues.

Verification and receipts

• Per-call execution receipts. Every tool call an agent makes is stamped with a typed receipt, visible on the run detail page — what was attempted, what happened, and when.
• Runs settle on evidence, not self-report. A run can only settle as complete when each side-effecting claim is backed by a receipt checked against ground truth — an email an agent says it sent must match a real delivery event, or the run is flagged instead of quietly marked done.
• Deploys don't drop work.In-flight runs survive platform deploys: a rolling restart requeues interrupted runs for re-dispatch, with guards against double execution, so a scheduled run doesn't silently vanish because the platform shipped an update mid-flight.

Security and tenancy (what's handled for you)

• Tenant isolation is enforced at the database with Row-Level Security and deny-all defaults, plus query-layer scoping by organization — the two checks compound.
• Inbound webhooks from external services are cryptographically signature-verified.
• Secrets (vendor OAuth tokens, model BYOK keys, signing keys, and integration credentials) are encrypted at rest.
• Single sign-on (SAML 2.0 / OIDC) is available for enterprise organizations.
• Self-healing with disclosure. Transient platform issues are retried and repaired automatically; when the same issue keeps recurring, you get exactly one deduplicated FYI in your approvals queue instead of silence or a flood.

For the full security posture — encryption, audit cadence, subprocessors, and responsible disclosure — see the Security page.