Persistent memory and cross-session learning for AI coding assistants (hosted remote MCP).
ContextStream MCP Server (io.contextstream/mcp)
ContextStream MCP Server is a hosted remote MCP that provides persistent memory and cross-session learning for AI coding assistants. It supports semantic code search and “team context” for Claude Code, Cursor, VS Code Copilot, Windsurf, and other MCP-compatible assistants.
🛠️ Key Features
Persistent memory
Cross-session learning
Semantic code search
Team context
MCP-compatible integration for AI coding assistants
🚀 Use Cases
Enable coding assistants to reuse context across sessions
Improve code retrieval via semantic code search
Provide shared team context to supported AI coding tools
⚡ Developer Benefits
Works with multiple AI coding assistants: Claude Code, Cursor, VS Code Copilot, Windsurf
MCP compatibility for “every MCP-compatible AI coding assistant”
⚠️ Limitations
Readme excerpt only reports benchmark information (no additional operational constraints described)
Performance claim is referenced as “90.0% on LongMemEval-S” with partial benchmark context shown
Your agent knows code. Give it the decisions behind yours.
Connect your code, docs, and conversations so your agents can find the right
files, recall saved decisions, and build on past work across sessions and tools.
The numbers, their sources, and what they do not claim are under Benchmarks.
Start free with MCP
10,000 monthly credits. No credit card required.
Works with Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI, Qwen Code,
Kimi Code, Muse Code, ZCode, Zed, and more MCP clients, on
Claude, GPT, Gemini, Kimi, GLM, Qwen, and other models.
Create your account or sign in during MCP onboarding. No separate website signup needed.
Use ContextStream to find the files relevant to my next change. Cite the
sources and retrieve any saved project decisions that should guide the work.
Then ask it to save a real decision and its reason. Start a new session and
retrieve that decision without explaining it again. Indexing supplies code
context; it cannot recover every undocumented decision.
Benchmarks
Each benchmark runs through the production API, is registered before its results
are read, and is published with its method, confidence interval, and losses.
Benchmark
ContextStream
Compared with
LongMemEval-S, all 500 questions
96.8% (484/500); 96.0% on the half held out from tuning
Above Mastra (94.87%) and supermemory (85.4%) on the same official GPT-4o judge. Level with agentmemory's self-reported 96.2%, which is inside our 95% interval
LoCoMo, all 1,540 questions
95.7% (1474/1540); 95.8% held out; 97.8% on the audited answer key
Above every system in the EverMemOS paper's same-judge table: EverMemOS 93.05%, Zep 85.22%, MemOS 80.76%, MemU 66.67%, Mem0 64.20%
Code search, 177 judged developer queries
98.77% best answer in the top 10; 100% a useful file in the top 10
Embedding-only search 72.84% and ripgrep 42.59% on the same queries and judgments
Agent tasks, 8 tasks × 3 trials
23/24 passed (96%); 9/9 on tasks that need team knowledge
14/24 (58%) and 0/9 with the repository alone
61 ms median server-side search time (127 ms p95) in the October 27 production
run. End to end it was 223 ms, because the client ran in a different region from the
one serving the workspace. Your latency depends on your distance from that region.
Competitor figures are as each vendor or the EverMemOS paper published them; we did
not re-run them, and readers and judges differ across rows. The code-search baselines
come from the August 1 run on the same queries. The agent-task result uses one
repository and one agent model, so treat it as directional.
Results, methodology, and limitations.
ContextStudio
For people ContextStream Create offers one-prompt making to, the server adds one tool,
studio: ask for a film, a product explainer filmed from your connected product, an
image, sound or music in plain words, and ContextStudio's producer plans the pieces from
what your team knows (brand kit, docs, decisions, connected products). There are no models
to pick.
Brand work by name. Brand images (social sets in each network's size, ad sets, hero
and product shots, posters and flyers with print PDFs, email headers, brand graphics) and
brand sounds (sonic logos, music beds, loops, effect packs, voice-overs, podcast intros
and outros): pass form, with sizes (up to 6, such as instagram-post or
linkedin-post) or a voice, and the plan keeps to it.
Plan, then price.plan returns the pieces and one price ("up to N credits, likely
about M") and spends nothing. Pass what your agent knows that the workspace may not (the
diff it just made, release notes, the doc it has open) as context; the producer reads
it next to the team's own context.
Your go-ahead.make spends at most the plan's price, and only after you agree: a
client that can ask you (MCP elicitation, or a stateless input_required result) asks
you itself; otherwise the agent passes the price you accepted.
Following it.status reports each piece while it waits (with
notifications/progress where the call asks for them); a product film takes 13 to 21
minutes. result shows images inline and links video and sound (signed links valid
about 15 minutes), with each piece's placard. Brand work comes as files, each with its
own link: every size as PNG, JPG and WebP with a ZIP of the set, a print PDF, or WAV and
MP3 for the web with a broadcast master and stems, plus alt text, a caption and the
checks it was held to.
Revising.revise starts from a made image or sound: a variation (more like it), a
revision (it, changed as you say) or a resize (an image in new sizes). Also stop,
library, brand and products.
Email, where Create offers it. Campaigns sent through your team's own Resend account.
email_draft writes one in your brand from a prompt ("announce our winter sale to
customers") and sends nothing; email_audiences, email_campaigns, email_test (to your
own address or Resend's test addresses), email_cancel and email_report (a campaign's
delivery, opens, clicks, bounces and complaints, or the workspace's trends). email_send
goes only with your go: the client asks you when it can; otherwise the agent shows you the
campaign and its recipient count and passes the count you agreed to.
email_preview gives a link to share that shows the email in Gmail, Outlook and Apple Mail
mockups (desktop, tablet and phone, light and dark, images on or off), with the inbox line
and what to check. email_revise changes a draft from plain words ("shorter subject") or
sets its subject and preview line; email_duplicate makes a draft copy. email_contacts
lists, adds (with where they agreed to hear from you) or unsubscribes contacts; erasing one
can't be undone, so it is the delete tool's (studio.email_contact_erase), after your go;
email_unsubscribes shows the suppression list with why and when, and takes an address off
it only with your go. Resend is connected in the studio, never through an agent: the tool
refuses anything shaped like a Resend key.
Connector. The hosted path /studio serves only this tool, for Claude.ai or ChatGPT
connectors, with OAuth tokens bound to that resource (read:studio, write:studio).
Data controls
Scoped access. Traceable sources. Configurable capture.
Indexing sends eligible source contents to ContextStream for hosted search.
Transcript saving and local Git metadata capture are on by default and can be
turned off. Review data handling and controls before
connecting sensitive projects.
Other installation options and troubleshooting
npm (Node.js 20+)
bash
npx -y @contextstream/mcp-server@latest setup
The npm launcher downloads the native Rust binary for your platform and verifies
its SHA-256 checksum. Pin an exact package version for production automation.
Hosted MCP without a local process
For clients supporting Streamable HTTP and OAuth, the hosted endpoint is:
Supported clients lists one-line add commands for Claude
Code, Codex, Gemini CLI, Qwen Code, Copilot CLI, and Droid, and one-click
install for VS Code. Use the MCP documentation
for other clients and stdio alternatives. A hosted connection alone does not
sync your local checkout.
Preview or diagnose setup
bash
contextstream-mcp setup --dry-run
contextstream-mcp doctor
contextstream-mcp doctor --repair
Scripted and CI setup
For automation, wrap the install in pipefail so it fails when the download
fails, instead of reporting success after installing nothing:
A hook for Claude Code and Cursor that checks the packages an install command names
against their public registry before the command runs. It covers npm, pnpm, yarn, bun,
npx, pip, uv, uvx, pipx, poetry and cargo, including claude mcp add ... -- npx ....
It denies a package the registry does not have, naming the popular package it
looks like. An agent may have invented the name, and anyone can register it.
It asks before a barely downloaded lookalike of a popular package.
It allows everything else, including whenever a lookup fails or takes too long.
It skips projects and commands that use a private registry (.npmrc, pip.conf, uv or
poetry sources, --registry, --index-url). CONTEXTSTREAM_PACKAGE_BOUNCER=off
turns it off.
Claude Code: add this entry to PreToolUse in ~/.claude/settings.json, next to the
hooks setup installed. update-hooks leaves it alone.
This repository contains the MIT-licensed Rust MCP server, client, editor setup,
and release tooling. The hosted backend is separate and is not included here.
See the architecture, release integrity,
and latest release.
Use the pinned Rust toolchain. For contribution checks and commit sign-off, read
CONTRIBUTING.md. Report vulnerabilities using
SECURITY.md. See LICENSE, NOTICE, and
GOVERNANCE.md for licensing, trademarks, and project governance.
Deep project learning starts off for new accounts. Interactive setup offers it
once per account (or once when terms change) and remembers if you decline; plain
contextstream-mcp setup --yes preserves your current choice.
bash
contextstream-mcp configure --account-learning status # show current consent state
contextstream-mcp configure --account-learning on # open privacy page in browser (or print URL)
contextstream-mcp configure --account-learning off # withdraw consent and queue deletion
--deep-learning is accepted as an alias for --account-learning. Only a person
can grant consent by ticking the unchecked box on the signed-in privacy page
(https://contextstream.io/dashboard/account/privacy); agents and API keys cannot
consent for you. Doctor and init show status only. See
data handling and controls.
Install
Remote endpoint
Streamable HTTP
Hosted server - connect over the network, no local install.