Capabilities
Audience: Engineers and integration leads exploring what the MCP server can do.
What you will accomplish: Navigate the full capability surface and understand how risk tiers, gates, and audit apply to each tool and service.
Prerequisites: Read Architecture first — you need the component model to reason about capabilities.
Estimated time: ~5 minutes (this page); ~30 minutes for the full section.
Where to start
New to the system? Follow this reading order:
- MCP Surface — Learn what tools, resources, and prompts the agent can invoke
- Workflow Engine — Understand how multi-step processes run with gates and crash recovery
- Core Services — Deep-dive into deadline engine, QC, document generation, and extraction
- Connector Layer — See how vendor APIs are abstracted behind ports
- Domain Packs — See how the engine is configured for a practice (immigration is the reference pack)
MCP Surface
Tools, resources, and prompts that the AI agent can invoke — all self-describing with Pydantic schemas.
Workflow Engine
Durable state-machine orchestration with human-in-the-loop gates, idempotency, crash recovery, and compensation.
Core Services
Deadline engine, QC verification, document generation, and data extraction — reusable across all workflows.
Connector Layer
Ports-and-adapters architecture — swap Clio → MyCase by changing one file, never touching core logic.
Domain Packs
The domain-agnostic seam — swap the practice (terminology, case types, rules, confidentiality policy) without touching the engine. Immigration is the reference pack.
Risk tiers
Every tool is classified by risk. The engine enforces the appropriate gate automatically.
| Tier | Description | Gate required? | Examples |
|---|---|---|---|
read | No side effects | No | matter.get, contact.search, deadline.compute |
write (confirm) | Reversible, single confirmation | Yes — one-click | matter.create, document.generate, form.prefill |
gated (human) | External or irreversible | Yes — explicit approval, default-on | email.send, document.route to external |
The gated (human) tier cannot be disabled per-workflow. This is a safety invariant, not a configuration option. Only read tools run without human involvement.
Capability matrix
| Capability | Status | Risk tier | Gate channel | Idempotent |
|---|---|---|---|---|
Client intake (intake.run) | Ready | Composite (confirm) | MCP / web / email | Yes |
| Status-update emails | Ready | Write (confirm) → Gated (send) | MCP / web / email | Yes |
| Document generation | Ready | Write (confirm) | MCP / web | Yes |
| Document routing | Ready | Write (confirm) / Gated (external) | MCP / web | Yes |
| Deadline computation | Ready | Read | None | N/A |
| Deadline scheduling | Ready | Write (confirm) | MCP / web | Yes |
| QC verification | Ready | Read (gate input) | None | N/A |
| Data extraction | Ready | Read | None | N/A |
| Contact upsert | Ready | Write (confirm) | MCP / web | Yes |
| Matter create | Ready | Write (confirm) | MCP / web | Yes |
What "ready" means
- Code complete: Implementation passes all unit, property-based, and integration-lite tests
- Documented: Workflow spec exists at
docs/workflows/<name>.md - CI-verified: Schema drift check passes; tool schema matches code
- Tested: Reference adapter enables full-fidelity testing without live vendor credentials
- Auditable: Every action writes a hash-chained record before returning
What is intentionally stubbed
| Component | Status | Reason |
|---|---|---|
| Vendor adapters (Clio, Lawmatics, Graph, SharePoint) | Stub | Awaiting firm vendor confirmation (PRD §11 Q1) |
| Production secret store (Vault / KMS) | Stub | Hosting decision pending (PRD §11 Q3) |
| PDF rendering engine (LibreOffice / WeasyPrint) | Stub | Engine confirmation pending |
| LLM provider integration | Mock only | Data residency policy pending (ASSUMPTION V-010) |
| Postgres-backed RunStore | In-memory only | Production migration ~3 days, tracked in roadmap |
Phase availability
| Phase | Timeline | Capabilities unlocked |
|---|---|---|
| 0 — Foundations | Complete | Domain model, MCP scaffold, audit log, read-only connectors, docs framework |
| 1 — Wave 1 | In progress | Intake workflow, status-update drafts, deadline engine, human gates |
| 2 — Wave 2 | Planned | Document generation, routing, extraction write paths, form pre-population |
| 3 — Wave 3 | Planned | Full QC suite, reconciliation, connector SDK, compliance certification |
Success metrics (current baseline)
| Metric | Baseline | Target |
|---|---|---|
| Test pass rate | 424/424 (100%) | Maintain ≥ 95% |
| Missed critical deadlines | N/A | 0 |
| Status updates sent on time | N/A | > 95% within SLA |
| QC defects reaching client | N/A | > 80% reduction |
| Connector onboarding time | N/A | < 1 week per new adapter |
Next steps
- MCP Surface — Complete tool/resource/prompt catalog
- Workflow Engine — State machines, gates, idempotency
- Core Services — Deadline engine, QC, document generation
- Connector Layer — Ports, adapters, and webhook ingestion
- Domain Packs — Configure the engine for any practice
Last updated: 2026-06-25