Carrier-native MCP server for the Guidewire InsuranceSuite. Ask Claude underwriter / claims / billing questions in operator vocabulary; the harness gates writes.
Links: GitHub · Live Architecture · v0.1.1 Release · Changelog
Insurance carriers run on Guidewire InsuranceSuite (PolicyCenter, ClaimCenter, BillingCenter). Underwriters, claims adjusters, and billing operators spend hours navigating the UI to answer questions like "what submissions are waiting on me?" or "why did we lose this account?" The Cloud API exists but requires developer intervention — operators can't self-serve.
A Claude Code plugin that exposes Guidewire as carrier-vocabulary MCP tools. Operators ask questions in their language (find-submissions-waiting-on-me), not API verbs (GET /job/v1/jobs). A governance harness gates all writes via plan → policy → approval → execute → audit → rollback. The hash-chained audit trail + carrier vocabulary are the durable moats.
| Who | What | Where | When | Why |
|---|---|---|---|---|
| Underwriters, claims adjusters, billing operators | Ask Claude natural-language carrier questions | Claude Code / Claude Desktop | Real-time during workflows | Reduce context-switching, enable self-service |
| Carrier IT / DevOps | Configure per-tenant profiles | 9-YAML profile directory | One-time setup | Customize LOBs, typelists, field aliases per carrier |
| Compliance / Audit | Review hash-chained audit trail | Postgres audit store | Post-hoc | Immutable record of all AI-assisted operations |
| Layer | Technology | Purpose |
|---|---|---|
| Agent Host | Claude Code, Claude Desktop, Anthropic API | Natural language interface |
| MCP Servers | servers/*-mcp/ (policycenter shipped; claim/billing planned) |
Tool catalog per InsuranceSuite module |
| Harness | packages/harness/ |
Gates writes via plan/policy/approval/execute/audit/rollback |
| Foundation | 7 packages (schemas, observability, auth, audit, client, mcp-runtime, harness) | Shared infrastructure |
| External | Guidewire Cloud API, Hub OAuth, Postgres, OTel, Sentry | Carrier systems + observability |
- Carrier vocabulary, not API verbs — Tool names are the question an operator would actually ask
- No mocks — Real Guidewire Cloud endpoints from day one
- Three execution modes —
read_only,draft_only,approved_executeper tool via profile - Hash-chained audit — Tamper-resistant via three Postgres roles (writer/reader/verifier)
- Per-tenant profiles — 9-YAML customization for LOBs, typelists, field aliases, approval matrix
Guidewire MCP for Claude v0.1.1 is a production-grade Claude Code plugin shipping:
- 7 E1 foundation packages
- 5 E2 read-only PolicyCenter tools
- E3 harness skeleton with approval workflow
- E4 profile loader scaffold
133 tests pass across 8 workspaces. Live architecture diagram at guidewire-mcp.intentsolutions.io.
v0.1.1 is a same-day install-path patch on top of v0.1.0 — fixes a silent-success failure mode in the prepare hook.
┌─────────────────────────────────────────────────────────────┐
│ Agent Host (Claude Code / Claude Desktop / Anthropic API) │
└─────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ MCP Servers (servers/*-mcp/) │
│ └── policycenter-mcp (5 read-only tools) │
└─────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ Harness (packages/harness/) │
│ └── plan → policy → approval → execute → audit → rollback │
└─────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ E1 Foundation Packages │
│ schemas │ observability │ auth │ audit │ client │ runtime │
└─────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────┐
│ External Systems │
│ Guidewire Cloud API │ Hub OAuth │ Postgres │ OTel │ Sentry │
└─────────────────────────────────────────────────────────────┘
# Install via Claude Code
/plugin install jeremylongshore/guidewire-mcp-for-claude
# Or develop locally
pnpm install # prepare hook builds all
pnpm -r build # rebuild all 8 workspaces
pnpm -r test # 133 tests
pnpm lint # biome check
node servers/policycenter-mcp/dist/cli.js --profile profiles/oss-demo| Tool | Purpose |
|---|---|
find-submissions-waiting-on-me |
Assigned-to-me queue |
show-policies-for-this-insured |
Cross-LOB policy lookup |
summarize-this-submission |
Full submission narrative |
did-we-lose-this-account |
Non-renewal / cancellation history |
pull-this-submission |
Single submission detail |
| Gate | Status |
|---|---|
| Tests (133) | ✓ Passing |
| Lint (Biome) | ✓ Clean |
| Typecheck | ✓ Passing |
| Build | ✓ All 8 workspaces |
| CI | ✓ 7 required checks |
| Branch protection | ✓ Ruleset enforced |
Apache-2.0
Jeremy Longshore — Intent Solutions · intentsolutions.io · jeremylongshore.com
All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
0.1.1 — 2026-05-06 — install-path patch (prepare hook)
Same-day patch on top of v0.1.0. Fixes the silent-success failure mode in the plug-and-play install path surfaced during the v0.1.0 release smoke-test.
preparehook no longer swallows build failures via|| true(#96). Ifpnpm -r buildfails duringpnpm installon an end-user's machine, the install now fails loudly with the underlying error instead of silently succeeding and leaving a non-functionaldist/cli.js. The silent-success failure mode was the worst kind: plugin installs cleanly, MCP server fails to register, user has no error message to act on. Surfaced during v0.1.0 release smoke-test (clone fresh to /tmp,pnpm install, verifydist/cli.jsproduced — test passed with current dependencies, but the silent-failure mode was unacceptable to leave in place).
0.1.0 — 2026-05-06 — E1 foundation + E2 read-only tools + E3 harness skeleton
First software release. E1 foundation packages, E2 PolicyCenter MCP with 5 read-only tools, and E3 harness skeleton are built and tested. 133 tests pass across 8 workspaces. Live architecture diagram at guidewire-mcp.intentsolutions.io.
E1 Foundation (7 packages):
@intentsolutions/guidewire-schemas— Zod schemas + TS contracts@intentsolutions/guidewire-observability— OTel + pino + Sentry factory@intentsolutions/guidewire-auth— Hub OAuth + JWT propagation@intentsolutions/guidewire-audit— Postgres + hash-chain audit store@intentsolutions/guidewire-client— undici Cloud API client with two-key idempotency@intentsolutions/guidewire-mcp-runtime— MCP SDK wrapper (stdio + HTTP transports)@intentsolutions/guidewire-harness— plan/policy/approval/execute/audit/rollback pipeline (#80)
E2 PolicyCenter MCP (5 read-only tools):
find-submissions-waiting-on-me— assigned-to-me queueshow-policies-for-this-insured— cross-LOB policy lookupsummarize-this-submission— full submission narrativedid-we-lose-this-account— non-renewal / cancellation historypull-this-submission— single submission detail
E3 Harness skeleton:
- Harness library + CLI skeleton — plan/policy/approval/execute/audit/rollback pipeline (#80)
- Postgres-backed ApprovalSink for approval workflow (#92)
- Boot-path error translator + CLI wiring (#93)
- BAA carve enforcement at boot when
lob_class:health(#87) oauthScope+idempotency.prunedplumbed through audit chain (#85)- Approvals table + grants in 0001_init.sql (#86)
E4 Profile loader scaffold:
- On-disk profile loader —
--profile <path>boots policycenter-mcp against a real tenant (#75) - 9-YAML schema: auth, roles, lob, typelists, custom-entities, field-aliases, approval-matrix, pii-policy, events
Infrastructure + tooling:
- Ship as Claude Code plugin —
/plugin install jeremylongshore/guidewire-mcp-for-claude(#76) - Karate contract testing for Cloud API layer (D-022) (#79)
- Testcontainers role-separation enforcement + pg-store TIMESTAMPTZ fix (#94)
- Live architecture diagram at guidewire-mcp.intentsolutions.io (#53)
- 5th project-level specialist:
guidewire-reference-librarian(#23) - 11-auditor staffed panel — security, MCP safety, harness contract, etc. (#52)
- README badges (CI, license, MCP-compatible, built-with-Claude) (#24)
- Workflow phase guards — CI no-ops without package.json, release is workflow_dispatch only
Documentation:
- 05-TECHNICAL-SPEC.md content (7.4k words) (#41)
- 04-USER-JOURNEY.md content (5.4k words) (#42)
- 01-BUSINESS-CASE.md content (3.0k words) (#39)
- 07-ROADMAP with E2.5 aggregate-query sub-epic (#43)
- Comprehensive CLAUDE.md rewrite reflecting current state (#77, #84)
- Public Guidewire docs reference (librarian KB) with PartnerConnect sandbox + Swagger UI (#54)
- Moved site from demo.intentsolutions.io to guidewire-mcp.intentsolutions.io (#53)
- CI/release workflows customized to pnpm + Biome + Vitest + audit-harness gates
- Main branch ruleset now enforced with 6 required status checks + 1 review (#90)
- CI workflow startup — dropped secrets ref from job-level if: condition (#88)
- Lint baseline from 43 → 0 — Biome ignore for Karate JS, auto-fix harness package (#89)
- Mobile-responsive architecture page — removed 1024px floor, added breakpoints (#71)
- Diagram layout — clean lane layout, separate desktop + mobile views (#72)
- Diagram mermaid syntax — frontmatter + escape primaryObject.id (#50, #55, #70)
- OWNER-bypass claim corrected — explicit bypass_actors required in ruleset (#91)
- Audit DB role separation enforced via testcontainers —
audit_writercannot UPDATE/DELETE (#94)
v0.0.1 — 2026-05-04 — paperwork foundation
Auto-cut by the release.yml workflow on initial push to main.
Marker for the paperwork-foundation milestone, NOT a software
release. Subsequent releases will be manual via workflow_dispatch
once code begins shipping in E1+.
- Repo scaffold + beads + Claude integration + git hooks
- 14 GH issues mirroring 14 beads via
bd-sync(12 epics + Phase 0 sub-beads + E11+ marketplace epic) - Phase 0 design inputs in
000-docs/:001-DR-RES-research-report.md,002-DR-CRIT-personas.md,003-DR-ARCH-oss-cowork.md,004-DR-DEC-architecture-decisions.md 000-docs/blueprint/— 12-section master blueprint scaffold (00 master index, 01-08 skeletons, 09 diagram placeholder, 10-AAR/ conventions, audits/ panel scaffold for 11 auditors).claude/agents/— 4 dual-mode specialist agent definitions (mcp-safety-reviewer,carrier-vocabulary-curator,guidewire-api-archaeologist,harness-runtime-architect)- Standard governance set: LICENSE (Apache-2.0), CHANGELOG, CODE_OF_CONDUCT, CONTRIBUTING, SECURITY, SUPPORT, .editorconfig, .gitattributes, GitHub issue/PR templates, CODEOWNERS, FUNDING, dependabot, default CI/release workflows
/validate-consistencypost-GW-1.1 audit ran clean after fix PR #17