Oceans LLM

Appearance

Overview and Features ​

See also: Documentation Home

Why Oceans LLM? ​

Oceans LLM gives teams one governed gateway for AI traffic: route requests across providers, manage identity and service access, control spend, observe usage, and expose MCP tools without asking every client to solve those problems separately.

Features ​

Access, Identity, and Permissions ​

  • SSO-ready admin access: Support for OIDC/OAuth SSO, including GitHub OAuth and configured OIDC providers,.

  • First-class users, teams, and service accounts: Model all identities separately, with service-account credentials, ownership, model grants, and budgets that remain independent from team membership.

    • Budgets are intentionly dropped for teams, avoiding spend abuse and promoting fairness.

  • Scoped API keys and model access: Gateway API keys for users or service accounts providing smoother on-boarding, granular model access, and credential revocation.

Routing and Client Experience ​

  • Provider routing behind one API surface: Let callers use OpenAI-compatible, Anthropic Messages-compatible, Responses, and embeddings endpoints while admins manage aliases, provider routes, capability gates, and compatibility behavior centrally.

  • Server-generated client setup: Generate client configuration snippets for supported model harnesses so users can connect tools like Claude Code, Codex, Opencode, and Pi with fewer manual setup mistakes.

  • Tags for attribution and reconciliation: Attach bounded request tags and identity tags so usage can be mapped back to cost centers, workloads, applications, environments, or existing internal company specific metadata.

  • Extensive Spend controls: Set budgets for users, service accounts, and model specific budgets by entity. Enabling access to frontier, costly models without risking runaway spend and complete loss of service.

Observability and Audit ​

  • Durable usage accounting: Record request usage into a ledger with pricing provenance, effective-dated rates, and clear priced, unpriced, and usage_missing states instead of relying on approximate billing.

  • FOCUS-oriented FinOps export: Export best-effort FOCUS-compatible CSV data for LLM usage so FinOps teams can ingest costs into existing tooling without maintaining brittle custom pipelines.

  • Request observability and audit detail: Inspect request logs with owner context, model resolution, provider attempts, latency, token usage, tags, payload policy, truncation state, and request correlation.

  • Agent harness adoption tracking: Understand which coding-agent clients, such as Claude Code, Opencode, Pi, Gemini CLI, Copilot CLI, and GitHub Copilot, are sending traffic through the gateway.

MCP Governance ​

  • Central MCP registry and access control: Register Streamable HTTP MCP servers once, discover tools, and expose either aggregate /mcp search/describe/call flows or direct per-server proxy routes.

  • Scoped MCP Tool Sets: Decompose multiple MCPs into smaller user specific toolsets to avoid tool overload and reduce risk of unintended tool execution.

  • MCP credential separation: Keep Oceans API keys separate from upstream MCP credentials, with gateway-managed and principal-bound credential bindings for safer tool execution.

  • MCP invocation logs: Audit individual tool calls with request correlation, owner context, server/tool identity, policy result, latency, status, and sanitized argument/result payload state.

Deployment and Operations ​

  • Deployment-shaped operations: Run locally with lightweight storage, deploy with PostgreSQL and Helm, configure secrets through environment references, and export OTLP traces and metrics to existing observability stacks.