Give your AI agent cross-session memory backed by Salesforce Data Cloud.
agent-context-sdk connects an agent you already run to the Salesforce
Context Service (Beacon). It does two things:
- Ingests each conversation turn — user messages, agent responses, reasoning, tool calls and results — so the service can build up context.
- Recalls that context on later turns and sessions, injecting it into the
agent's prompt and exposing it as tools (
recall_customer_context, context packs).
It ships with adapters for LangGraph, the Claude Agent SDK, Google ADK, the OpenAI Agents SDK and Microsoft Foundry, all in one package.
pip install agent-context-sdkRequires Python 3.10+. All five adapters are included. The framework SDKs
themselves (langgraph, claude-agent-sdk, google-adk, openai-agents,
agent-framework-foundry) are not bundled — you already run the one your
agent uses. If it's missing, the adapter tells you exactly which package to
install.
1. Authenticate with a Salesforce Connected App. For an end user signing in from a local tool:
from agentic_context import UserContextAuth
auth = UserContextAuth.login(
consumer_key="3MVG...", # Connected App consumer key
login_url="login.salesforce.com", # login host, no scheme
aid="your_agent_id", # agent id registered with the Context Service
)For a backend service account, use ContextAuth (JWT bearer) instead — see
the Quickstart.
2. Wrap your agent with its adapter. For example, the Claude Agent SDK:
from claude_agent_sdk import ClaudeAgentOptions, query
from agentic_context import ClaudeAdapter
adapter = ClaudeAdapter(agent_id="your_agent_id", owner_id=auth.user_id, auth=auth)
options = adapter.apply_to(ClaudeAgentOptions(model="<claude-model>"), session_id="ses-123")
async for message in query(prompt="Hi, I need help with my order", options=options):
...Every turn is now ingested, and prior context is recalled automatically. The other adapters follow the same pattern:
from agentic_context import (
ClaudeAdapter, GoogleAdkAdapter, LangGraphAdapter, FoundryAdapter, OpenAIAdapter,
)Imports are lazy — importing one adapter only loads that framework's SDK.
If you don't use one of the supported frameworks, AgentContextClient gives
your own agent loop the same ingest and recall without an adapter. Your code
opens the conversation and ingests each message; the model gets two mid-turn
tools (ace_get_context, ace_get_context_pack) via tool_specs() /
call_tool(). It is a preview API and not semver-committed — see the
AgentContextClient guide.
| Guide | What's in it |
|---|---|
| Quickstart | Credentials and per-framework wiring for all five adapters |
| Developer guide | Full per-adapter guide: prerequisites, constructors, troubleshooting |
| Advanced | Backend-hosted login, token storage, service accounts, static tokens |
| AgentContextClient | Adapter-free client for custom agent loops (preview) |
| Examples | Runnable demos |
| Changelog | Release notes |
┌─────────────────────────────────────────────┐
│ Framework adapter layer │
│ LangGraph / Claude / ADK / OpenAI / Foundry │
│ Turns agent events into Record objects │
└────────────────────┬────────────────────────┘
│ list[Record]
▼
┌─────────────────────────────────────────────┐
│ SessionManager / AsyncSessionManager │
│ Creates the conversation + session lazily │
│ Retries transient failures │
│ Async path never blocks the agent │
└────────────────────┬────────────────────────┘
│ HTTPS
▼
┌─────────────────────────────────────────────┐
│ Context Service │
└─────────────────────────────────────────────┘
Each Record carries a stable local event_id (for logs and callbacks). It
is not sent to the service, so re-sending a turn writes duplicate entries.
pip install -e ".[dev]" # SDK plus the framework SDKs the adapter tests need
pytest # core tests
pytest adapters/<fw>/tests # one adapter's tests (adk, claude, foundry, langgraph, openai)
ruff check src tests # lint
mypy src/agentic_context # type-checkLayout: the core lives in src/agentic_context/; each adapter lives in
adapters/<fw>/src/ and is built into the same wheel. For adapter design
rules, see docs/Design.md.
Report bugs and request features through GitHub Issues. This project does not accept external code contributions; see CONTRIBUTING.md.
Do not report security issues in GitHub Issues. See SECURITY.md.
Licensed under the Salesforce Public Code License — see LICENSE.txt.