Best Weights & Biases MCP Server Alternatives 2026

Knowledge graph-based persistent memory system. Store and retrieve contextual information.

Free and open-source by Anthropic. Knowledge graph stored locally — no external service costs.

a single structured-reasoning tool that lets a model plan, revise and branch its own chain of thought instead of answering in one shot. Published by Anthropic as part of the official modelcontextprotocol/servers monorepo (89,000+ stars, actively maintained), it exposes exactly one tool — sequential_thinking — and that tool is the whole product. Each call carries a `thought` string plus bookkeeping fields: `thoughtNumber`, `totalThoughts`, and `nextThoughtNeeded`, which the model flips to false when it is done. The interesting fields are the optional ones. `isRevision` and `revisesThought` let the model go back and correct an earlier step rather than plowing ahead on a bad assumption; `branchFromThought` and `branchId` let it fork into an alternative line of reasoning and carry both forward; `needsMoreThoughts` lets it extend past its own original estimate when a problem turns out to be deeper than it looked. In practice you never call the tool by hand. You connect the server to an MCP host and ask a question that deserves more than one pass — plan a PostgreSQL 14 to 16 migration and revise if downtime exceeds five minutes, work out why a deploy only fails in production, compare three architectures and branch when an assumption breaks. You can tell it is working when the host inspector shows repeated sequential_thinking calls with a rising `thoughtNumber` rather than a single response. Install with `npx -y @modelcontextprotocol/server-sequential-thinking` — note the hyphenated package name, which differs from both the `sequentialthinking` directory in the repo and the Docker image `mcp/sequentialthinking`, a mismatch that breaks a lot of copied configs. A Docker image is published alongside the npm package, and the README carries one-click VS Code install buttons for both transports. Set `DISABLE_THOUGHT_LOGGING=true` if you do not want every thought written to the server log.

Free and open-source by Anthropic. Runs locally with no external API dependencies.

Exa's official MCP server connects AI assistants to a search engine purpose-built for AI, using neural embeddings to match on meaning rather than keywords so agents get clean, ready-to-use content instead of a page of blue links to re-parse. The default tool set covers web_search_exa for quick topical lookups and web_search_advanced_exa for full control over domains, date ranges, and content filters, plus specialized tools for code_search (searching real-world code and GitHub), company_research (building company profiles, competitor lists, and financials), crawling/web_fetch (pulling clean content from a specific URL), people_search and linkedin_search (public professional-profile lookups), and deep_researcher_start/check for long-running multi-step research tasks backed by Exa's Research API. The server is hosted at https://mcp.exa.ai/mcp — no local process to run — and connects via one-line setup in Cursor, VS Code, Claude Code, Claude Desktop (available as a native Connector), Codex, OpenCode, Windsurf, and Antigravity, authenticated with an EXA_API_KEY from the Exa dashboard. Tool exposure is tunable per client via a ?tools= query parameter on the endpoint URL, letting teams ship narrow, purpose-built configurations (e.g. company-research-only or LinkedIn-only agents) instead of exposing the full surface, and Exa ships ready-made Claude Skills/agent definitions for common patterns like company research and people search with built-in query-variation and token-isolation guidance.

The MCP server is free and open-source. Exa: Free tier with 1,000 searches/mo. Pro plans available for higher volume. See official pricing.

OpenAI Codex is a lightweight coding agent that runs in your terminal, and it sits on both sides of the Model Context Protocol — which is exactly the confusion the "codex mcp" search hides. As a client, Codex connects to other MCP servers you configure under `[mcp_servers]` in `~/.codex/config.toml`, and the `codex mcp` subcommand manages those configured launchers. As a server, Codex exposes its own experimental MCP interface: run `codex mcp-server` (alias `codex-mcp-server`) and another MCP client can drive a local Codex engine over stdio using JSON-RPC 2.0. That server interface is genuinely useful and genuinely experimental — the docs state it is subject to change without notice — and it publishes methods to manage threads, turns, accounts, config and approvals: v2 RPCs like `thread/start`, `thread/resume`, `thread/fork`, `turn/start`, `turn/steer`, `turn/interrupt`, `account/login/start`, `config/read`/`config/value/write`, `model/list`, plus v1 compatibility calls (`getConversationSummary`, `getAuthStatus`, `gitDiffToRemote`, fuzzy file search) and streaming `codex/event/*` notifications. Approvals flow the other way as server-to-client requests (`applyPatchApproval`, `execCommandApproval`), so a host UI can gate patch and exec actions. Install the CLI with `npm install -g @openai/codex` (or Homebrew), then start the server with `codex mcp-server`; a quick inspection UI is `npx @modelcontextprotocol/inspector codex mcp-server`. Built in Rust, Apache-2.0 licensed, and by far the most-starred entry in this space at 100,000+ stars. Use this when you want an AI client to orchestrate a real Codex coding session rather than reimplement one.

This MCP server is free and open-source. Check the GitHub repository for details.

ClickHouse MCP Server is ClickHouse's official MCP server (ClickHouse/mcp-clickhouse) that connects Claude, Cursor, and other MCP clients to a ClickHouse cluster for fast analytical querying over natural language. Its primary tool, run_query, executes arbitrary SQL against your cluster in read-only mode by default (CLICKHOUSE_ALLOW_WRITE_ACCESS=false) so an AI assistant can explore tables, aggregate billions of rows, and answer analytics questions without risk of mutating data — writes can be enabled explicitly when needed. Companion tools list databases and tables and return schema metadata (including the full create_table_query, with an option to omit per-column detail for lighter responses). A second tool set embeds chDB, ClickHouse's in-process engine, via run_chdb_select_query, letting the assistant query files, URLs, and external databases directly without an ETL step (enabled with the optional mcp-clickhouse[chdb] extra). Destructive statements are gated a second time: even with writes enabled, DROP and TRUNCATE require CLICKHOUSE_ALLOW_DROP=true as well. The server supports both stdio and HTTP/SSE transports; on HTTP/SSE authentication is required rather than optional — startup fails unless a static bearer token (CLICKHOUSE_MCP_AUTH_TOKEN), a FastMCP OAuth/OIDC provider (Azure Entra, Google, GitHub, WorkOS via FASTMCP_SERVER_AUTH), or an explicit local-development opt-out (CLICKHOUSE_MCP_AUTH_DISABLED) is configured. Connection is configured through CLICKHOUSE_HOST, CLICKHOUSE_PORT, CLICKHOUSE_USER, and CLICKHOUSE_PASSWORD, with ClickHouse Cloud, self-hosted, and the public SQL playground all supported.

The MCP server is free and open-source. ClickHouse Cloud: Free trial available. Pay-as-you-go pricing. Self-hosted ClickHouse is free and open-source.

The Milvus MCP Server (zilliztech/mcp-server-milvus) is Zilliz's official Model Context Protocol bridge into Milvus, the open-source vector database used for RAG pipelines, semantic search, and large-scale embedding retrieval. It gives an AI agent direct, structured access to a running Milvus instance rather than requiring hand-written client code: milvus_vector_search runs similarity search over embeddings, milvus_text_search performs full-text search, and milvus_hybrid_search and milvus_text_similarity_search combine both approaches (the latter requires Milvus 2.6.0+ with a server-side embedding function configured). Beyond search, the server covers full collection lifecycle management — milvus_list_collections, milvus_create_collection (quick-setup or custom schema), milvus_get_collection_info, milvus_load_collection/milvus_release_collection for memory management, plus milvus_query for filter-expression lookups, milvus_insert_data, and milvus_delete_entities for direct data mutation. It supports three transport modes: stdio (default, for local Claude Desktop/Cursor use), SSE for multi-client HTTP access, and Streamable HTTP with an optional stateless flag for production deployments without session persistence. Install with `uv run src/mcp_server_milvus/server.py --milvus-uri http://localhost:19530` after cloning the repo, or point it at any local or remote Milvus cluster by URI — Python 3.10+ and the uv package manager are the only prerequisites. Works with Claude Desktop, Cursor, and any other MCP-compliant client.

Free and open-source MCP server. Milvus is open-source. Zilliz Cloud (managed): Free tier available, paid plans for production use.

Embeddings, vector search, document storage, and full-text search with the open-source AI application database.

Free and open-source MCP server. Chroma is open-source and free to self-host. Chroma Cloud currently in development.

The Sentry MCP Server is Sentry's official Model Context Protocol integration, purpose-built for human-in-the-loop coding agents like Claude Code, Cursor, and Windsurf. Rather than exposing every Sentry API endpoint, it focuses tightly on developer debugging workflows: searching and triaging issues, pulling stack traces and event details, inspecting performance traces, and querying project/team/org metadata in natural language. The primary deployment is a hosted remote MCP server at mcp.sentry.dev, built on Cloudflare's remote-MCP infrastructure, so most users connect with zero local setup — just add the remote URL to their client. For self-hosted Sentry instances or local development, a stdio transport is also available via npx @sentry/mcp-server, authenticated with a Sentry User Auth Token scoped to org:read, project:read, project:write, team:read, team:write, and event:write. AI-powered search tools (search_events, search_issues) translate natural-language queries into Sentry's query syntax, but require a configured LLM provider (OpenAI, Azure OpenAI, Anthropic, or OpenRouter) — all other tools work without one. Claude Code users can also install it as a plugin (claude plugin install sentry-mcp@sentry-mcp) for automatic subagent delegation whenever a conversation touches Sentry errors, issues, or traces. This turns "why did this deploy break in production" into a direct conversational debugging session instead of tab-switching into the Sentry dashboard.

The MCP server is free and open-source. Sentry: Developer tier (free, 5K errors/mo). Team: $26/mo. Business: $80/mo. Enterprise: Custom.

The Datadog MCP Server is Datadog's official, vendor-hosted Model Context Protocol endpoint — not a package. Each Datadog site has its own URL of the form https://mcp.<your-site>/api/unstable/mcp-server/mcp (US1: mcp.datadoghq.com, EU1: mcp.datadoghq.eu), and OAuth 2.0 is the recommended way in; a Personal or Service Access Token as an Authorization bearer header is the documented fallback for CI, with DD_API_KEY plus DD_APPLICATION_KEY headers as a third option. Datadog ships first-party client integrations rather than expecting hand-written config: a Claude Code plugin (/plugin install datadog@claude-plugins-official, then /ddsetup and /ddtoolsets), a Claude connector from the Connectors Directory, plugins for Cursor, VS Code/Copilot, JetBrains and OpenCode, and a ChatGPT app in Preview for US1. Tools are grouped into toolsets selected with a ?toolsets= query parameter, and only `core` — logs, metrics, traces, dashboards, monitors, incidents, hosts, services, events, notebooks — loads by default; two dozen more cover alerting, DBM, DDSQL, RUM, profiling, security, Kubernetes, error tracking, feature flags, cost management and data observability, with apm, cases, code-exec and remote-actions in Preview and excluded from toolsets=all. Access requires the mcp_read or mcp_write role permission in addition to the normal resource permission, which is why a working connection can still return no data. Limits at time of writing are 50 requests per 10 seconds of tool-call burst and 50,000 tool calls per month, and the server is not GovCloud compatible.

The MCP server is free and open-source. Datadog: Free tier (5 hosts). Pro: $15/host/mo. Enterprise: $23/host/mo. Additional products priced separately.

The official Grafana MCP server connects Claude and other AI assistants directly to your Grafana instance and its surrounding observability ecosystem, turning natural-language questions into dashboard lookups, incident investigations, and datasource queries. Dashboard tools cover search, retrieval, JSONPath-scoped property extraction, patch-based editing, and per-panel query/datasource introspection, with context-window-aware helpers like get_dashboard_summary so an agent never has to pull a full multi-megabyte dashboard JSON just to answer a simple question. Query tools speak PromQL against Prometheus (including histogram-percentile helpers), LogQL against Loki, and native query languages for InfluxDB, ClickHouse, CloudWatch, Graphite, Athena, Snowflake, Elasticsearch/OpenSearch, and Quickwit datasources — most gated behind opt-in --enabled-tools flags to keep the default tool surface lean. It also wraps Grafana Incident for creating and updating incidents, Sift for automated error-pattern and slow-request investigations, full alerting CRUD (rules, contact points, notification policies) across Grafana-managed and external Alertmanager sources, Grafana OnCall schedule/shift/alert-group management, RBAC-gated admin tools for teams/users/roles, deeplink generation so the LLM never has to guess a dashboard URL, annotations, snapshots, PNG rendering via the Grafana Image Renderer, and provisioning-repo validation for git-sync workflows. Authentication is a Grafana service account token (Editor role, or granular RBAC scopes) passed as GRAFANA_SERVICE_ACCOUNT_TOKEN alongside GRAFANA_URL, and every tool category can be individually disabled to control context-window usage. On install, the recommended route is uvx: `uvx mcp-grafana` pulls the PyPI package mcp-grafana, which is published by Grafana Labs from this same repository — so despite the server being written in Go, the copy-paste command most Claude Desktop and Cursor configs use is a Python-tooling one, not a binary download. The alternatives are `go install github.com/grafana/mcp-grafana/cmd/mcp-grafana@latest` for a real binary, or the grafana/mcp-grafana container — `-t stdio` for local clients, or the default HTTP mode on :8000 (add `-t streamable-http`) with MCP_GRAFANA_SERVER_TOKEN set to authenticate callers when you expose it.

The MCP server is free and open-source. Grafana Cloud: Free tier (10K metrics, 50GB logs). Pro: $29/mo. Advanced: $299/mo. Self-hosted is free.