What it does

ContextStream keeps project context outside any single agent session, so a new coding agent does not have to be briefed from scratch. It resolves the authorised workspace and project, searches eligible indexed source for paths and lines, and can inspect code relationships where the index supports them. Separately, it stores and retrieves decisions and their reasons, preferences, constraints, lessons and handoffs. The agent uses that material to propose a next step, and you check the cited evidence before acting. The site is candid that it does not guarantee agents will use the context correctly, and that it does not build or validate the application being coded. Search is hosted: eligible file contents are sent for indexing even if the MCP helper runs locally, so the page tells you to review repository policy and exclusions before ingest. Setup connects a supported client, account, workspace, project and index, and the MCP helper installs with a curl script. The site also publishes benchmark results for memory accuracy and code recall. The page lays out other ways to carry context, such as briefing agents yourself, keeping notes in the repo or relying on an agent's built-in memory, and suggests testing one small, checkable handoff before adopting it.

Capabilities

Hosted code search with paths, lines and code relationships
Saved decisions, lessons, guardrails, handoffs and runbooks
Works with coding agents such as Cursor, Claude Code and Codex through MCP
Returned sources can be checked before you approve a plan
Privacy and indexing settings to review before ingest

When to use it

Several agents or tools work on one project and keep losing earlier decisions
You want handoffs and lessons to carry over between sessions
You are allowed to send the repository to a hosted index