Ask Sage · Palantir AIP · IBM watsonx · Databricks — Combined & Corrected Reference

Merged from two independent analyses, fact-checked against primary sources, and reconciled. Prepared July 13, 2026.

Provenance key used throughout: [A] = carried from the original attached analysis and verified accurate · [C] = added or corrected from this review's research · [A→C] = present in the original but corrected here. Full change log in Section 9.

Claude models, Claude Code / the Claude Agent SDK, and enterprise AI platforms operate at three different layers. The central error to avoid is treating them as competing, substitutable products:

• Claude models provide reasoning, language generation, coding, and tool selection.
• Claude Code and the Claude Agent SDK provide an agent harness: context assembly, tool execution, permissions, agent loops, memory, hooks, subagents, skills, plugins, and IDE interaction.
• Enterprise AI platforms provide governed data, identity, workflow execution, semantic models, lineage, deployment, evaluation, compliance, and integration with operational systems. [A]

Using a Claude model does not mean using Claude Code's architecture. Among the platforms reviewed:

• Ask Sage has the closest direct compatibility with the actual Claude Code client — confirmed: its Anthropic-compatible endpoint runs the real Claude Code CLI, VS Code extension, and Claude Cowork unmodified, including DoD PKI cert support for Army GenAI. [A, verified]
• Palantir has gone furthest in combining the Claude Agent SDK, MCP, enterprise permissions, operational data, and actions — confirmed via Foundry's Anthropic-compatible proxy endpoints and Ontology MCP's support for headless Anthropic SDK agents. [A, verified]
• Databricks has an equally direct Claude Code relationship, not merely a model-serving one: Agent Bricks lists the Claude Code SDK as a first-class supported harness, and Omnigent — Databricks' open-source "harness of harnesses" — composes Claude Code, Codex, and custom agents under one governed runtime. This reclassifies Databricks from "not documented" to Direct. [C — corrects the original analysis]
• IBM has the strongest proprietary alternative for hybrid enterprise integration, legacy modernization, decision services, and mainframe-heavy environments, via IBM Bob and watsonx Orchestrate as a multi-vendor agentic control plane. [A]
• Claude Code itself remains the strongest of these for repository-centric coding and configurable local agent behavior. [A]

One fact changes the practical calculus for a DoD environment specifically, and it is addressed nowhere in the original analysis — see Section 2.

New in this version — absent from the original attached analysis, and material to any recommendation that assumes Claude is a freely approved model inside DoD contracts and classified networks.

In late February 2026, DoD designated Anthropic a "supply chain risk" — a designation typically reserved for foreign adversaries — following a dispute over whether the military could use Claude for "all lawful purposes," which Anthropic said could authorize autonomous weapons systems or domestic mass surveillance. President Trump directed federal agencies to cease using Anthropic technology. On May 1, 2026, DoD announced classified-network (IL6/IL7) AI partnerships with eight vendors — SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, AWS, and Oracle — pointedly excluding Anthropic. Anthropic sued; a California injunction preserves its ability to serve civilian agencies, but the Pentagon exclusion stands for DoD contracts while litigation continues. Reporting since indicates informal talks have reopened, but as of this writing the exclusion is active and contested.

Practical implication: recommendations below that route through "Claude via the approved AIP model service" or "Claude via Ask Sage" remain architecturally correct — both platforms genuinely support Claude as described — but whether Claude is a procurable, authorized model for a specific DoD contract or classified network is a live policy question independent of the technical architecture. Two consequences:

• Architect any DoD-facing build so the model provider is swappable behind your MCP/tool abstraction layer — Claude, or another accredited model, should sit behind the same harness without a rebuild.
• For unclassified CUI/FOUO work (most day-to-day DoD FM analysis), Claude via FedRAMP-High pathways may remain more available than classified-network deployments — confirm current status with your ISSM/AO before committing engineering time, since this is actively litigated and changing.

Claude Code is more than an interface to an LLM: it is an agentic coding tool that reads a codebase, edits files, executes commands, and interacts with development tools; the Claude Agent SDK exposes the same tools, agent loop, and context-management mechanisms for custom agents. [A]

Claude Code assembles context from source files, CLAUDE.md, local auto-memory (indexed by MEMORY.md, capped at the first 200 lines / 25KB per session load), active IDE files, command output, MCP resources/tools, agent skills, and subagent results. This is flexible, session-oriented project knowledge — not the same thing as an enterprise knowledge graph, governed data catalog, or semantic SQL layer. [A, verified]

• Skills: reusable SKILL.md-centered instruction packages that load on demand — relevant DoD FM examples include audit evidence validation, USSGL/SFIS mapping checks, NFR root-cause analysis, and reconciliation procedures. [A]
• Hooks: deterministic controls at lifecycle events (block an unsafe command, run a scan post-edit, require a test before completion, log to an audit system).
• Plugins: package skills, hooks, subagents, MCP servers, and LSP servers into one distributable, versioned unit.
• Subagents / agent teams: context-isolated specialist workers with restricted tools, run in foreground or background.
• LSP: plugins can bundle language-server diagnostics so Claude sees IDE errors/warnings directly; a documented "safe mode" flag disables CLAUDE.md, plugins, skills, hooks, and MCP servers for troubleshooting. [A, core claim verified; LSP/auto-memory inclusion in safe mode specifically unconfirmed]

These matter much less for large-scale financial transaction analysis than data lineage, semantic modeling, SQL execution, access control, and evidence provenance. [A]

An AI harness is the system around the model that determines whether it becomes a useful enterprise agent or merely a chatbot. A mature harness includes twelve functional components: [A]

Claude Code and the Agent SDK supply much of this for coding-oriented agents. Palantir, IBM, and Databricks supply alternative enterprise harnesses. Ask Sage can either supply its own Deep Agent/Agent Builder environment or serve as the secure model endpoint behind the actual Claude Code harness. [A]

Classification terms: Direct = the platform explicitly implements or supports the Claude mechanism. Client-provided = the mechanism works because the real Claude Code client is being used against the platform's endpoint. Analogous = the platform has a proprietary equivalent, not format-compatible. Not documented = no public documentation establishes compatibility. [A]

Ask Sage's Anthropic-compatible endpoint is fully documented to run the real Claude Code CLI and VS Code extension with minimal configuration (point ANTHROPIC_BASE_URL at Ask Sage), including DoD PKI certificate support for Army GenAI environments, and the same pattern extends to Claude Cowork. Because the real client is doing the work, Ask Sage does not need to reimplement CLAUDE.md, skills, hooks, plugins, subagents, or LSP — it supplies model routing, authentication, government connectivity, and inference underneath the unmodified Anthropic client. [A, verified]

Ask Sage's own native platform separately includes datasets, persona/prompt templates, ~27 plugins and agents, Deep Agent, Agent Builder, workbooks/canvases, and API access — but these should not be treated as format-compatible with Anthropic Agent Skills, Claude plugins, or CLAUDE.md. [A]

• Access to 150+ commercial models behind one controlled, government-accredited interface
• Fastest path to Claude Code inside a controlled DoD/IL5/IL6 environment without building a model gateway
• CUI-capable, zero-trust, label-based access control; edge/air-gapped deployment option
• Lowest-engineering path from chat to API-based applications for non-developer analysts

Ask Sage should not be treated as the authoritative enterprise data or audit system. It is weaker than Palantir or Databricks on enterprise semantic modeling of financial objects, transaction-scale reconciliation, column-level lineage, distributed SQL/Spark computation, and evidence-chain management. For DoD audit, position it as a secure model and analyst-assistance layer — not the Universe of Transactions, reconciliation engine, or evidence repository. [A]

• Level 1: Claude as an AIP model, via Azure, AWS Bedrock, Google Vertex, or native enrollment, subject to region/environment availability.
• Level 2: Claude Code as an external Foundry client — confirmed. Foundry provides LLM-provider-compatible proxy endpoints (Anthropic, OpenAI format) specifically so tools like Claude Code can run while Foundry adds rate limiting, zero data retention, and usage tracking. Under this pattern Claude Code supplies CLAUDE.md, skills, plugins, hooks, subagents, and IDE integration; Palantir supplies governed data, Ontology, models, APIs, permissions, and enterprise tools; MCP connects the two.
• Level 3: Claude Agent SDK inside Palantir-hosted agents — confirmed. Ontology MCP explicitly documents integration with headless agent frameworks including the "Anthropic SDK," alongside Google ADK, Microsoft Agent Framework, and OpenAI SDK.

Ontology MCP (OMCP) exposes object types, action types, and query functions as governed MCP tools for agents that need to safely read, write (via predefined actions only), and query production ontology data; access is scoped through application restrictions and standard Foundry permissions. Palantir MCP is a separate, developer-facing server (80+ tools across roughly 14 categories) for building Foundry resources — datasets, ontology, code repos, OSDK apps — and explicitly cannot write production ontology data. This separation is far more appropriate for high-consequence financial management than giving an agent unrestricted database or shell access. [A, verified]

AI FDE (Palantir's own enterprise harness) acts under the current user's identity and permissions — if the user cannot create a repository, edit an object type, or execute an action, neither can AI FDE. Palantir replaces most Claude Code constructs with its own:

[A]

• Operational data integration and business-object modeling — an Ontology can represent financial statement line items, USSGL accounts, appropriations, journal vouchers, obligations, disbursements, feeder systems, supporting documents, audit samples, controls, NFRs, and corrective actions as first-class, related objects
• Fine-grained authorization/markings, human-in-the-loop actions, case management, cross-system lineage
• Structurally satisfies EU AI Act Article 13-style auditability with less custom configuration than competing platforms

The Ontology requires careful design and sustained governance; poor semantic modeling produces a sophisticated interface over unreliable data. Proprietary constructs create vendor dependence. Large-scale raw computation is more naturally a Spark-oriented job where Databricks is already the data platform. [A]

IBM and Anthropic announced a strategic partnership (October 2025) to infuse Claude into IBM's software portfolio with enterprise security, governance, and cost controls. IBM is primarily adopting Claude models, not replacing its architecture with Claude Code's. Its harnesses remain watsonx Orchestrate, IBM Agent Development Kit, IBM Agent Catalog, watsonx.data, watsonx.governance, IBM Bob, and IBM's decision/automation products. Orchestrate now functions as an agentic control plane governing agents from IBM's own Granite models, Anthropic's Claude, OpenAI's GPT, Mistral, and custom builds under one consistent policy layer — IBM's real differentiator is governing agents it didn't build, at scale, across vendors. [A, extended]

Bob (IBM's coding agent, replacing the earlier Watsonx Code Assistant) covers architecture, coding, testing, security, and modernization for Java, IBM i, and mainframe environments, dynamically routing tasks across Claude, Mistral, and Granite by accuracy/performance/cost — roughly 80,000 IBM employees reportedly use it internally with a cited 45% productivity gain. Public documentation does not establish Bob as using CLAUDE.md, Anthropic Agent Skills, Claude plugin packaging, hooks, LSP plugins, or Claude Code's exact subagent configuration — it is its own harness and development model. [A, extended]

• Hybrid/on-premises architecture; Java, COBOL, RPG, and mainframe modernization
• Business rules and decision services — a deterministic entitlement or policy calculation should remain in a rules engine, with Claude explaining the policy or gathering inputs, not improvising the final number
• $11B Confluent acquisition brings Kafka/Flink real-time streaming into watsonx.data; Sovereign Core embeds policy at the infrastructure runtime level

Less attractive when the goal is the exact Claude Code developer experience, lightweight greenfield agent development, Spark-scale data engineering, or a Palantir-style operational ontology. Its breadth (Orchestrate, watsonx.data, watsonx.ai, watsonx.governance, Bob) can itself create architectural complexity. [A]

This section corrects the original analysis, which understated Databricks' Claude Code relationship based on documentation that predates the June 2026 Data + AI Summit.

Databricks supports Anthropic through Mosaic AI Model Serving with centralized credential management and AI Gateway controls, and its Agent Framework supports custom agents, RAG, tool-calling, and multi-agent systems with MLflow tracing/evaluation. At the June 2026 Data + AI Summit, Databricks explicitly expanded Agent Bricks — its developer agent platform (100,000+ agents built, reportedly processing over a quadrillion tokens/year) — to support the Claude Code SDK as a first-class harness alongside LangGraph, Agno, CrewAI, and OpenAI Agent SDKs. Databricks also released Omnigent, an open-source "harness of harnesses" that explicitly composes Claude Code, Codex, and custom agents in one workflow, with contextual policies enforced at runtime through Unity AI Gateway and every session traced in MLflow. This is Direct adoption of the Claude Agent SDK/Claude Code harness — not merely hosting the underlying model. [C]

Unity Catalog is the governance backbone — it enforces access controls, tracks data/AI lineage down to the column level, records activity for audit, and now governs tables, files, functions, models, agents, and MCP servers together (MCP was added directly to Unity Catalog). Genie Ontology continuously learns business semantics (fiscal year definitions, org hierarchy, term definitions) so agents don't rebuild context every call, and Genie Agents convert natural-language questions into governed SQL. This is generally more reliable for financial metrics than placing business definitions only in a system prompt. [A, extended with C]

• Terabyte/petabyte-scale processing, Spark and distributed SQL, data engineering pipelines, ML feature development
• The most direct interoperability with the actual Claude Code harness of any platform reviewed, via Agent Bricks and Omnigent
• Unity AI Gateway as a single governance/cost/guardrail layer across models, agents, MCP servers, and Skills

Less naturally oriented toward operational object/action modeling, case-based remediation workflows, and human-approval processes tied directly to business objects — those can be built, but typically require additional applications or workflow systems on top. GovCloud/FedRAMP High coverage is expanding but currently behind Ask Sage and Palantir's classified-environment accreditation timeline. [A, extended]

Examples: policy analysis, audit report review, correspondence drafting, NFR summarization, control narrative generation.

Recommended architecture: secure model gateway (Ask Sage or an enterprise model endpoint) → document ingestion and metadata extraction → hybrid keyword+vector retrieval → reranking with document-level access control → Claude Agent SDK or platform-native agent → citation and evidence display → human review → evaluation against a curated Q&A set.

Measure retrieval correctness, citation accuracy, whether superseded guidance was excluded, and whether the system appropriately declined when evidence was insufficient — not how fluent the answer sounds. Ask Sage is the strong default here; Palantir becomes more valuable once documents must tie to operational objects, transactions, controls, owners, or actions. [A]

Examples: budget execution, transaction reconciliation, GL analysis, abnormal journal review, obligation-to-invoice matching.

Recommended architecture: governed structured data layer → data contracts and quality rules → semantic layer defining measures and accounting concepts → deterministic SQL/code tools → an MCP interface exposing only curated queries and actions → the LLM for planning, tool selection, and explanation → reproducible result storage → human approval for consequential actions.

The LLM should never calculate critical financial totals from thousands of records in its own context window: SQL calculates the balance, Spark performs the large join, a rules engine tests compliance, a statistical model flags anomalies — the model chooses the approved tool and explains results. Choose Palantir when data must connect to business objects, cases, and actions; Databricks when the center of gravity is SQL/Spark/ML/scale; IBM when logic spans legacy/mainframe systems; Ask Sage as the analyst-facing language layer on top. [A]

For millions or billions of transactions, never send raw records to Claude. Pattern: raw source systems → distributed ingestion/standardization → data-quality and reconciliation pipelines → curated transaction/balance/exception tables → semantic measures and approved query functions → MCP tools → Claude or another LLM → explanation, prioritization, and workflow initiation.

Focus areas: partitioning, incremental processing, data skew, source-to-target reconciliation, schema evolution, lineage, deterministic business rules, cost per million transactions, and false-positive management. Databricks generally has the best native foundation for this computation; Palantir can consume the curated results and convert them into operational cases and approved actions. A common mistake is using an agent to replace the Spark/SQL layer — the agent should orchestrate the computation, not perform it in natural language. [A]

GAO's May 2026 testimony (GAO-26-109115) confirms DoD's revised audit approach shifts from decentralized response and reliance on internal controls toward centralized coordination (the new Joint Task Force-Audit), a focus on material line items, and "manual testing of large samples and using artificial intelligence tools, as needed" — explicitly increased emphasis on technology including AI, with evidence supporting material account balances taking priority over remediating underlying control deficiencies for at least the next two years. [A→C: verified against the primary GAO source and paraphrase tightened]

The target should not be "an audit chatbot." It should be an audit evidence and validation architecture with AI assistance, in seven layers: [A]

Adopt directly: the Claude Agent SDK for the agent loop; MCP for tool/data interfaces; Agent Skills or a similar version-controlled procedure format; a CLAUDE.md-like repository policy file; subagents for isolated specialist tasks; hooks for deterministic safety/quality controls; sandboxed execution; explicit permissions and tool approvals; complete tracing and replay; automated evaluation. [A]

Keep vendor-neutral: store critical business logic outside model-specific prompts — SQL for calculations, Python/Spark for transformations, MCP/OpenAPI for tools, versioned schemas, OpenTelemetry-compatible traces, model-independent evaluation datasets, external identity/authorization, external evidence storage. The model should be replaceable without rebuilding the system — this is now a functional requirement, not just good hygiene, given Section 2. [A→C]

Do not overinvest in LSP for audit applications — it is valuable for software development but not a priority for financial-audit users. For audit, invest instead in data lineage, query validation, evidence provenance, document-to-transaction linking, access controls, semantic definitions, reproducible calculations, and review/approval workflows. [A]

• Text-heavy secure assistant → prioritize Ask Sage plus Claude Code or the Claude Agent SDK.
• Data-heavy operational solution → prioritize Palantir AIP and the Ontology.
• Big-data analytical solution → prioritize Databricks with Unity Catalog, SQL/Spark, MLflow, and Agent Bricks/Omnigent if Claude Code-style harness portability matters.
• Legacy, hybrid, or mainframe-heavy environment → prioritize IBM watsonx and IBM Bob.
• DoD audit → the layered architecture in Section 7, with the model layer built swappable per Section 2.

For DoD audit specifically, the flow is:

Authoritative financial/business data → Reconciliation, lineage, and semantic controls → Palantir/Advana evidence and operational ontology → Governed MCP tools → Claude or another approved model → Analyst explanation, exception prioritization, and controlled workflow → Human review and immutable audit record

The strongest overall DoD audit design is not a single commercial AI product. It is a controlled combination of a trusted data and evidence plane, deterministic audit logic, governed tools, a constrained agent harness, an approved LLM (with the provider treated as swappable given the current DoD–Anthropic contract dispute), and human accountability. [A→C]

Non-negotiable principle: the AI may recommend, explain, classify, retrieve, and orchestrate. The governed data platform must calculate, reconcile, authorize, record, and prove. That separation will matter far more to audit success than which model or vendor is selected. [A]

Transparency on what changed between the original attached analysis and this combined version, and why.

Everything not listed in this table was checked against available primary sources during this review and found consistent with the original analysis. No other material factual errors were identified — the original document's sourcing discipline was, on the whole, unusually strong.