Connecting your data

Give agents access to external sources using context providers.

An agent is only as good as the context it has access to. Context providers give the agent live access to a source through two natural-language tools: query_<source> for reads and update_<source> for writes. A sub-agent behind each provider owns the source’s quirks, so the main agent’s context stays clean.

from agno.agent import Agent
from agno.context.web import ExaBackend, WebContextProvider
from agno.models.openai import OpenAIResponses

web = WebContextProvider(backend=ExaBackend(), model=OpenAIResponses(id="gpt-5.5"))

agent = Agent(
    model=OpenAIResponses(id="gpt-5.5"),
    tools=web.get_tools(),
    instructions=web.instructions() + "\nAlways cite URLs inline.",
)

agent.run("What's the latest pricing for Anthropic's models?")

Navigation over search

Context providers navigate the source at query time, the way a coding agent runs ls, grep, and cat. The state is always live and every citation is a path the user can open.

The common “company brain” approach ingests everything into a vector database for top-k retrieval. That works for static corpora. For live product data, the index goes stale and citations point at fragments.

Win	Why it matters for your product
Live state	The Slack message sent thirty seconds ago is queryable now
Real citations	Every reference is a path the user can open
Permissions stay put	The source enforces who can read what; the agent sees what its credentials see

Why a provider, not raw tools

Attaching a source’s twelve raw tools to the agent pollutes context and degrades routing. A provider collapses the source to one tool the agent reasons about.

Problem with raw tools	What a context provider does
Too many tools crowd the prompt	The agent sees one tool: query_slack
Overlapping scopes confuse routing	One sub-agent owns one scope; routing is one tool call
Intermediate results bury the task	Pagination and lookups stay inside the sub-agent

Sources

Source	Tools	Backed by
Web	query_web	Exa, Parallel, or an MCP search backend
Workspace	query_workspace	A filesystem path or repo, read-only
Database	query_<db>, update_<db>	Postgres, with a read engine and a schema-guarded write engine
Knowledge wiki	query_knowledge, update_knowledge	Filesystem or a Git repo for durable prose memory
Slack	query_slack	Slack API, read-only
Google Drive	query_gdrive	A service account, scoped to shared folders
MCP servers	query_mcp_<slug>	Any MCP server: Linear, GitHub, Notion

Read-only providers

Providers that can write take write=False, which removes the update tool so the agent cannot mutate the source under any prompt. A read-only source like the web has no write tool to begin with. This is how a style guide or a reference corpus stays immutable while still being queryable.

Next steps

Task	Guide
Reach users where they work	Interfaces
Persist what the agent learns	Sessions and memory

Developer Resources

• Context engineering
• Context providers cookbook
• Context providers in the runtime
• Scout: connect MCP servers