Substrate Skill for Claude

Butterbase Substrate

Substrate is an optional, per-user add-on backend. One substrate per Butterbase account, lazily provisioned on first use, queryable from local Claude Code (via MCP), from any opted-in Butterbase app (via ctx.substrate inside functions), and from external systems (via HTTP).

Four stores

| Store | What lives there | Tools |

|---|---|---|

| Entities | People, companies, projects you transact with | get_entity, find_entities |

| Business state | Numeric facts that change over time (MRR, headcount) | find_entities with type filter |

| Institutional memory | Decisions, commitments, learnings | search_memory |

| Action ledger | Append-only log of every proposed/executed action | list_outbox, propose_action history |

The propose → policy → execute → log loop

Every write goes through this loop. Agents NEVER touch the substrate database directly.

• Propose — propose_action with capability + args. Returns an action ID.
• Policy — substrate-core evaluates per-capability rules. Verdicts: auto_execute, require_approval, deny.
• Approval (if needed) — human calls approve_action(action_id) or reject_action(action_id). If yolo_mode=true for this user, approvals are auto-granted for capabilities marked yolo-safe.
• Execute + log — substrate-core runs the action in a transaction and appends to the ledger.

For external side effects (sending email, calling an API), substrate writes to the outbox instead of executing inline. The cron-scheduler drains it. Use list_outbox, retry_outbox, cancel_outbox to manage it.

Reads

• get_entity(id) — fetch by ID.
• find_entities(type, query, limit) — typed search; query is full-text + structured.
• search_memory(query, kind?) — semantic + keyword search across decisions/commitments/learnings.

When to use

• ✅ Founder copilot ("what did I decide about pricing last quarter?")
• ✅ AI agent that should remember a customer across conversations
• ✅ Any app where you want a single source of truth for entities the agent operates on
• ❌ Plain CRUD app (use regular Butterbase tables)
• ❌ Pure analytics / read-only dashboard (substrate is action-oriented)

Linking an app

When apps.substrate_user_id is set to the app owner's platform_user.id, functions in that app get ctx.substrate injected at cold start, with reader and proposer methods. Use /butterbase-skills:journey-substrate to enable this for an app.

API keys

For agent access from local Claude Code, generate a scope='both' API key via manage_auth_config action: "generate_service_key" with substrate_access: true. The same bb_sk_ key works on both app and substrate endpoints.

Anti-patterns

• ❌ Calling substrate MCP tools without first checking whether the user has substrate provisioned. find_entities returns 401 if there's no substrate — handle gracefully and suggest provisioning.
• ❌ Treating the action ledger as mutable. It's append-only; "undo" is a new compensating action, not a delete.
• ❌ Storing transient state in entities. Entities are durable nouns; use the app's runtime tables for ephemeral state.
• ❌ Calling upsert_entity without canonical_keys or primary_email and expecting dedup. Provide one or the other so substrate can match an existing row. Without them, every call mints a new entity.
• ❌ Calling update_entity for a partial update. It replaces attrs wholesale and will drop every key you didn't include. Use patch_entity (RFC 7396 merge-patch) instead.
• ❌ Storing entity IDs in app data and skipping alias resolution after merge_entities. The merged-away ID stops resolving to an entity; look up substrate.entity_aliases to find the survivor.