Intelligence without context is unreliable. Autonomy without context is chaos.
And context without trust becomes surveillance.
Give every coding agent persistent memory, semantic code search, and the right project context on every turn.
Works with Claude Code, Cursor, VS Code + Copilot, Windsurf, Cline, Roo Code, Kilo Code, Codex, OpenCode, Aider, Antigravity, and other MCP clients.
Install in 30 seconds · Benchmarks · Documentation · Start free
Your AI should not have to rediscover the repository, re-learn the architecture, or repeat last week's mistake every time a chat starts. ContextStream gives MCP clients a shared context layer that remembers decisions and lessons, searches code by meaning, maps dependencies, and restores the thread after compaction.
The server in this repository is the canonical, MIT-licensed Rust
implementation. The recommended editor transport is the hosted MCP at
https://mcp.contextstream.io/mcp; a small local helper keeps explicitly linked
checkouts in sync.
We publish the full method, confidence intervals, comparison caveats, and known losses—not just the winning number. The product-level figures below are from the published July 2026 report; the LongMemEval-S result was published June 14, 2026.
| Benchmark | ContextStream result | What was measured |
|---|---|---|
| LongMemEval-S | 90.0% | 450/500 correct on the full suite; ~115k-token multi-session histories; official pinned GPT-4o judge; gpt-5.5 reader with self-consistency k=3 (89.6% single-shot); Wilson 95% CI [87.1%, 92.3%] |
| Pinned-repo code search | 99.38% Recall@10 | 93.83% Primary@10 and 95 ms successful-response p50 over 177 judged-valid queries, averaged across two cold and two warm runs; 0/708 failed requests; no local checkout |
| Agentic project memory | 96% task success | Up from 58% without the memory layer; one repository, one agent model, three trials per cell, so this result is directional |
On the standard memory benchmark, 90.0% is statistically above supermemory's published 85.4% and statistically tied with Zep's published 90.2%. Those comparisons use each vendor's published result; they are not presented as a private same-run head-to-head.
Read the full benchmark report, methodology, per-category results, and losses →
The repository also includes Criterion suites for JSON-RPC parsing, tool
dispatch, response serialization, and stdio/HTTP transport primitives. Run
them locally with cargo bench --locked -p mcp-server.
Pick the command for your platform. Each installer downloads the current native release and launches the guided setup wizard.
curl -fsSL https://contextstream.io/scripts/setup.sh | bashNative releases are published for macOS Apple Silicon and Intel, plus Linux ARM64 and x64.
irm https://contextstream.io/scripts/setup.ps1 | iexThe Windows installer uses the native x64 binary and installs it in your user profile, so it does not need an administrator shell.
npx -y @contextstream/mcp-server@latest setupThe dependency-free npm launcher selects the matching release artifact, verifies its SHA-256 checksum, caches it by exact package version, and starts the Rust binary. It supports macOS ARM64/x64, Linux ARM64/x64, and Windows x64.
The wizard signs you in, detects supported editors, writes their hosted MCP configuration and managed rules, installs lifecycle hooks where supported, links the exact project folder you choose, starts indexing, and checks the result. You review the choices before it makes changes.
After setup, restart the editor and try:
Use ContextStream to summarize this repository, cite the files you relied on, and tell me one active decision or risk that should affect my next change.
Want to inspect the changes first? Every setup write can be previewed:
contextstream-mcp setup --dry-runAlready configured and troubleshooting? Run:
contextstream-mcp doctor --scope=all --only-configuredIf your MCP client supports Streamable HTTP and OAuth, connect directly—there is no local MCP process to keep running:
{
"mcpServers": {
"contextstream": {
"type": "http",
"url": "https://mcp.contextstream.io/mcp"
}
}
}Some clients use a different root key or config format, which is why the setup wizard is the easiest path. The hosted gateway is the default; the installed helper only syncs the local checkouts you explicitly link.
Clients that require stdio can use the verified npm launcher:
{
"mcpServers": {
"contextstream": {
"command": "npx",
"args": ["-y", "@contextstream/mcp-server@latest"],
"env": {
"CONTEXTSTREAM_API_KEY": "your-api-key"
}
}
}
}For production automation, pin an exact 1.x package version instead of
latest.
| Capability | What changes in practice |
|---|---|
| Grounded context | Each turn starts with the relevant workspace rules, prior decisions, lessons, and current project state. |
| Semantic code search | Find implementations by intent, with hybrid, keyword, pattern, exhaustive, and refactor-aware modes when precision matters. |
| Durable memory | Decisions, plans, tasks, runbooks, preferences, and session transcripts remain queryable across chats and tools. |
| Compaction survival | Important state is checkpointed before long conversations compact and restored afterward. |
| Dependency intelligence | Trace blast radius, dependencies, circular imports, complexity, and unused code without manually rebuilding the graph. |
| Agent handoffs | Package verified state and next steps into durable handoffs and portable ContextCapsules. |
| Team context | Bring shared workspace knowledge and supported GitHub, Slack, and Notion context into the same MCP surface. |
The default tool surface is grouped into focused domains—context, search,
memory, session, graph, entity, capsule, project, workspace,
skill, media, vcs, and more—so agents get broad capability without an
enormous tool-registration prompt.
The wizard currently detects and configures:
- Claude Code
- Cursor
- Windsurf
- VS Code + GitHub Copilot
- Cline
- Kilo Code
- Roo Code
- OpenAI Codex
- Aider
- Antigravity
- OpenCode
Anything else that speaks the Model Context Protocol can connect to the hosted URL or use the stdio launcher.
This repository contains the canonical Rust MCP server, protocol types, client and session layers, editor setup, hooks, public transport code, release automation, and the MongoDB-free acceleration foundation. It builds as seven workspace crates:
mcp-servermcp-toolsmcp-clientmcp-sessionmcp-typesmcp-model-registrymcp-acceleration-products
The hosted backend, credentials, production deployment configuration, and
premium provider implementations are not included. Public-build compatibility
traits remain no-op so the core protocol surface stays stable. See
docs/architecture.md for the boundary.
Setup shows a disclosure before changing editor configuration or indexing a
project. You choose the folder and workspace relationship; only matching source
files from a validated project can be indexed. Transcript capture, local Git
metadata capture, ignore rules, kill switches, retention behavior, and deletion
paths are documented in docs/data-handling.md.
Review that document before using setup --yes in automation. The CLI also
provides --dry-run for setup, repair, migration, and uninstall workflows.
The workspace uses the Rust toolchain pinned in rust-toolchain.toml.
cargo build --locked -p mcp-server --bin contextstream-mcp
cargo test --locked --workspace
cargo bench --locked -p mcp-serverTo build the hosted HTTP gateway foundation:
cargo build --locked --release -p mcp-server --bin contextstream-mcp \
--features remote-accelerationEvery release publishes native binaries, checksums.txt, version.json, and
an SPDX SBOM. The npm launcher downloads only from the immutable path for its
exact package version, verifies manifests and binary SHA-256, uses an atomic
versioned cache, and supports verified offline reuse with
CONTEXTSTREAM_MCP_OFFLINE=1.
See docs/release.md for the release contract and
the latest GitHub release
for artifacts.
All 1.x npm releases preserve these executable aliases:
mcp-servercontextstream-mcpcontextstream-hook(forwards to the Rusthooksubcommand)
Issues and pull requests are welcome. See CONTRIBUTING.md,
SECURITY.md, and GOVERNANCE.md. Contributions
must include the Developer Certificate of Origin sign-off (git commit -s).
The code is MIT licensed. ContextStream trademarks and the hosted-service boundary are described in NOTICE.
Stop teaching your coding agent the same project twice.
Install ContextStream ·
Start free ·
Read the docs
