Meko MCP server icon

Meko

by Meko

HIPAA CompliantSOC2 ReadyISO 27001 Ready
Analytics23 tools

Give AI agents persistent memory, shared knowledge and file storage on a hosted YugabyteDB layer. Anthropic lists 23 tools; Meko documents the same 23 by name, including four deletes. OAuth or an API key, identity-only scopes.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on .

Anthropic states this reflects the level of review a connector received, not a security audit.

Connect Meko via MCP

https://mcp.mekodata.ai/mcp

Works in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.

Meko Tools & Capabilities (23)

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

  • No application scopes exist. The RFC 9728 resource descriptor and the authorization server both advertise exactly openid, email and profile, and the 401 challenge repeats the same three. These are identity claims. Nothing at the OAuth grant separates reading a memory from deleting a datapack, and the boundary a reader would most want — between the routine write and the irreversible one — does not exist here. Meko instead enforces authorization by authenticated identity and datapack role grants, which is a real boundary but not one you choose at consent.
  • Tool-level scope was retired and is a breaking change. Meko removed the per-call scope argument (read, write, admin) in installer v2.1.0. A compatibility shim currently strips it from older clients and Meko says that shim will be removed. Custom integrations still sending it must stop.
  • One credential reaches every datapack you own. Keys are account-scoped, capped at 10 active per account, and expire 365 days after creation. Revocation is immediate and permanent.
  • Knowledge-base ingestion is not available over MCP. Cloud uploads go through the portal UI — PDF, TXT, MD, JSON or MP4, 5 MB each, 10 per batch. knowledgebase_search only queries the resulting index.
  • There are no MCP tools to clean up RAG state. Meko's skills documentation states there is no way over MCP to delete a vector index, remove a source from one, clear stuck queue entries or reset a failed pipeline, and that deleting a knowledge-base registration through the control plane does not remove the underlying index, source records or vector data. Its documented workaround is to create a new index under a different name and leave the stale one for an admin.
  • Memory extraction is best-effort and poor at structured data. Meko states automatic extraction can miss an individual fact, and that memory_add works badly for CSV rows, data dictionaries and schemas because columns, rows and field relationships get dropped. Its advice is to write one narrative summary rather than ingest row by row.
  • Automatic capture is broken or partial in three clients. Meko's v2.1.0 notes state Cursor's ordinary archive and new-chat workflows do not reliably emit the lifecycle events capture depends on; Codex transcript capture "isn't functional" because the parser does not understand Codex response_item records and can advance its watermark after extracting nothing; and Codex sessions currently derive a claude_code namespace rather than a codex one. Claude Desktop has no hooks at all, so capture there depends on the model choosing to call the tool, which Meko says Sonnet- and Opus-class models typically will not do without an always-on profile instruction.
  • No semantic search over conversation content. Finding a past conversation means browsing with conversation_list and inspecting candidates with conversation_get. conversation_search is hidden unless a deployment enables it.
  • We could not read tool schemas or safety annotations. The endpoint returned 401 to an anonymous request. Parameter names on this page come from Meko's documentation and its public skills repository, not from the server.
  • Meko's llms.txt contradicts its own reference on authentication, instructing agents to send an API_KEY header that the reference page says does not authenticate. Two other names in that file are not tools: its "Suggested first tools" list is accurate, but Meko's pricing page shows an agent-facing snippet calling knowledge_search, which is not a tool — the real name is knowledgebase_search.
  • Meko's pricing page contains a section addressed to AI models, headed "For AI agents reading this page" and giving a four-step integration procedure. We read it as data and report it here; we did not follow it as instruction. Treat vendor pages that instruct models as content to evaluate, not as direction.
  • No Python SDK exists. Meko states framework SDKs are under development and MCP is the integration path until they ship.

Frequently asked questions

Meko is a hosted memory and context layer for AI agents, built by YugabyteDB on a distributed PostgreSQL backend. Its documentation describes four stores inside a workspace it calls a datapack: extracted facts, saved conversation transcripts, an uploaded-document knowledge base, and content-addressed file artifacts. The MCP server is how an agent reads and writes all four.

Yes, through four named tools. Meko's reference documents conversation_delete as permanently deleting a conversation and all its messages, datapack_delete as irreversible, memory_delete_by_id as removing one memory, and memory_delete_all as a full reset for an agent. A datapack deletion takes its memory, conversations, knowledge base and artifacts with it.

Not on the 23-tool surface. Meko's documentation carries three db_run tools inside HTML comments, including a write-SQL one, but they are commented out of the rendered page and appear in no shipped tool list. Direct SQL is instead a separate PostgreSQL connection string, outside MCP entirely and outside anything an agent can reach.

Only identity claims. The server's RFC 9728 descriptor advertised openid, email and profile on 2026-08-23, and its authorization server advertises exactly the same three. None of them names an object or a verb, so the consent screen offers no way to grant searching memory without also granting deleting a datapack.

Yes. Meko's quick start states that API keys and the MCP connection are scoped to your account rather than to a single datapack, and that once connected your agent can see and act on every datapack you own. One endpoint serves them all; the agent picks one by passing a datapack ID on each call.

Yes, when you install the optional hooks. Meko's directory-review notes state that session and compaction hooks can send user prompts, assistant responses, tool-call summaries, tool-result summaries, working directory, git branch and session ID to Meko. Meko documents that installing only the skill files manually, without the hooks, avoids automatic capture.

They consume a metered quota rather than money. Meko's reference documents Free-tier caps of 1,000 conversations and 10,000 retrievals per month, and states that further calls are blocked when a quota is exceeded. No tool reads the remaining balance, so an agent cannot check its quota before spending it.

Because OAuth-only setups are on demand. Meko's troubleshooting states that if you authenticate with OAuth alone and skip the automatic installer, you must mention Meko in your prompt for the agent to call the tools, and it will not proactively save conversations or capture memories. The automated install adds skills and hooks that close that gap.

Sources

Use in Agentman

Connect once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.

Open in Agentman Studio

Server Info

Category
Analytics
Developer
Meko
Tools
23
Domain
mcp.mekodata.ai

Using Claude Desktop or another MCP client? Setup docs — the connection URL above works anywhere.