Database for AI Agents | TiDB Distributed SQL | TiDB

Expand instantly to match demand, then scale to zero when idle

• Autoscaling adapts to agent workloads automatically, from zero to peak
• Instant branching creates isolated environments in milliseconds
• Compute and storage scale independently, no capacity planning required

Query vectors, transactions, and analytics in a single call

• Join vector similarity results directly against transactional data
• One engine replaces your vector store, warehouse, and transaction layer
• No ETL, no middleware, fewer systems to build and maintain

-- One query. Vectors + transactions.
-- No separate vector DB needed.

SELECT
  a.agent_id, a.goal, a.status,
  m_top.relevance
FROM agent_tasks a
JOIN (
  SELECT id,
    vec_cosine_distance(
      embedding, '[0.12, 0.85, ...]'
    ) AS relevance
  FROM memory_store
  ORDER BY vec_cosine_distance(
    embedding, '[0.12, 0.85, ...]')
  LIMIT 100
) m_top ON a.context_id = m_top.id
WHERE a.status = 'active'
ORDER BY m_top.relevance
LIMIT 10;

Give every agent action a reliable, up-to-date view of the data

• Agent state stays ACID-consistent from one workflow step to the next
• Concurrent agents always read identical state, no matter which node they hit
• Multi-step workflows survive failover without losing progress or replaying actions

-- Agent step 3: update, then read back.
-- The read ALWAYS reflects the write.

BEGIN;

UPDATE agent_state SET
  step = 3,
  plan = JSON_SET(plan,
    '$.next_action', 'execute')
WHERE agent_id = 'a-7f3e'
  AND step = 2;  -- compare-and-swap

-- Immediate read: no stale state.
SELECT step, plan->>>'$.next_action'
  FROM agent_state
  WHERE agent_id = 'a-7f3e';
-- Returns: step=3, next_action='execute'
-- Even across distributed nodes.

COMMIT;

Run concurrent AI agents without interference or compromise

• Each agent gets dedicated throughput that other workloads can’t touch
• Tenant-scoped data separation keeps every agent’s state private
• No shared resources, cross-tenant contention, or coordination overhead

Build with the tools and syntax your agents and developers already know

• Standard SQL means agents and developers start productive on day one
• MCP, pytidb, LangChain, and LlamaIndex integrations out of the box
• Online DDLs and zero-downtime schema changes so shipping never stops

// Connect any AI agent to TiDB
// via the MCP server.

{
  "mcpServers": {
    "TiDB Cloud": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "https://tidb.link/mcp"
      ]
    }
  }
}

// Now your agent can:
// "Create a branch from prod"
// "Run SELECT * FROM agents"
// "Show me slow queries"

Expand instantly to match demand, then scale to zero when idle

• Autoscaling adapts to agent workloads automatically, from zero to peak
• Instant branching creates isolated environments in milliseconds
• Compute and storage scale independently, no capacity planning required

Query vectors, transactions, and analytics in a single call

• Join vector similarity results directly against transactional data
• One engine replaces your vector store, warehouse, and transaction layer
• No ETL, no middleware, fewer systems to build and maintain

-- One query. Vectors + transactions.
-- No separate vector DB needed.

SELECT
  a.agent_id, a.goal, a.status,
  m_top.relevance
FROM agent_tasks a
JOIN (
  SELECT id,
    vec_cosine_distance(
      embedding, '[0.12, 0.85, ...]'
    ) AS relevance
  FROM memory_store
  ORDER BY vec_cosine_distance(
    embedding, '[0.12, 0.85, ...]')
  LIMIT 100
) m_top ON a.context_id = m_top.id
WHERE a.status = 'active'
ORDER BY m_top.relevance
LIMIT 10;

Give every agent action a reliable, up-to-date view of the data

• Agent state stays ACID-consistent from one workflow step to the next
• Concurrent agents always read identical state, no matter which node they hit
• Multi-step workflows survive failover without losing progress or replaying actions

-- Agent step 3: update, then read back.
-- The read ALWAYS reflects the write.

BEGIN;

UPDATE agent_state SET
  step = 3,
  plan = JSON_SET(plan,
    '$.next_action', 'execute')
WHERE agent_id = 'a-7f3e'
  AND step = 2;  -- compare-and-swap

-- Immediate read: no stale state.
SELECT step, plan->>>'$.next_action'
  FROM agent_state
  WHERE agent_id = 'a-7f3e';
-- Returns: step=3, next_action='execute'
-- Even across distributed nodes.

COMMIT;

Run concurrent AI agents without interference or compromise

• Each agent gets dedicated throughput that other workloads can’t touch
• Tenant-scoped data separation keeps every agent’s state private
• No shared resources, cross-tenant contention, or coordination overhead

Build with the tools and syntax your agents and developers already know

• Standard SQL means agents and developers start productive on day one
• MCP, pytidb, LangChain, and LlamaIndex integrations out of the box
• Online DDLs and zero-downtime schema changes so shipping never stops

// Connect any AI agent to TiDB
// via the MCP server.

{
  "mcpServers": {
    "TiDB Cloud": {
      "command": "npx",
      "args": [
        "-y", "mcp-remote",
        "https://tidb.link/mcp"
      ]
    }
  }
}

// Now your agent can:
// "Create a branch from prod"
// "Run SELECT * FROM agents"
// "Show me slow queries"

Book a 30-Minute Meeting

Please enter your information in the form and we’ll contact you shortly.