ContextStream

ContextStream's production product is the MCP connection used from the assistant you already have. It is implemented by the Rust MCP server and connects Claude Code, Cursor, Codex, VS Code, and other supported clients to ContextStream's search, memory, and durable project context.

Setup takes one command. Afterwards, your existing assistant can do things it couldn't do reliably from an isolated chat—and it can decide when to reach for that project context instead of waiting for you to prescribe every lookup.

The decision from three weeks ago, not a fresh opinion.

Files and lines, from a pre-built index, without reading the repository first.

The exchanges and snapshots your connected assistants captured — including one you used in another editor.

“remember we're on Postgres, not MySQL”

Saved as durable context for your authorized connected clients.

“write the signing-key rotation runbook”

A maintained operating procedure with verification, rollback, and stop conditions.

The indexed image matched from its extracted text and visual context.

“handoff the remaining auth review to another agent”

A scoped transfer with current state, open questions, next steps, and evidence.

It can keep durable context.Approved plans, explicit decisions and lessons, and captured exchanges persist between sessions, assistants, and machines. A record saved in Cursor can ground Claude Code tomorrow.

It stops guessing where things are.Your repository is indexed, so finding code is a lookup instead of a search-and-read loop. That's faster and cheaper — fewer tokens spent reading files that turned out to be irrelevant.

You can share selected context when you collaborate.A shared decision someone else recorded is available to your assistant when the topic comes up. If you work alone, the same workspace and project scopes organize your own context across sessions and assistants. Personal records and uncaptured conversations remain outside shared context.

The intended loop is simple: you state the outcome and constraints; your assistant decides what evidence and prior work it needs; ContextStream supplies the relevant project context and keeps useful results available for later.

You should not need to say “search the code, now read the decisions, now check the lessons, now find the runbook.” For a task such as “fix the refresh race, preserve the public API, and show me the verification,” the assistant can orient to the project and retrieve those sources on its own initiative.

Before rotating a signing key, ask the assistant where verification happens and what depends on it. Have it read the decision that set the overlap window and the lesson from the last rotation. Save an executable plan and link the runbook. Set a reminder for when work must begin, not the expiry date.

If you stop halfway through, save the exact state. A later session can resume with the open task and failed verification. If someone else takes over, create a handoff; if their agent needs the supporting context, pair it with a narrow, expiring capsule.

You still speak in normal sentences throughout. The connection keeps “where is it?”, “why is it this way?”, “what do I do?”, “what remains?”, and “who owns it now?” as different kinds of project context so each question has a useful answer later.

The setup wizard manages Claude Code · Cursor · GitHub Copilot in VS Code · Windsurf · Cline · Kilo Code · Roo Code · Codex CLI · Aider · Antigravity · OpenCode CLI.

Your assistant connects to ContextStream through the hosted MCP endpoint. The installed Rust binary configures supported clients, generates their managed instructions, refreshes those instructions during the update flow, and supplies the machine-local source sync path needed for indexing. The normal editor connection remains hosted; the helper on your machine does not replace it with a second local product.

You do not need to learn internal operation names to use this. Setup writes instructions that teach each supported assistant when to search, recall, or save durable context, so ordinary use stays conversational.