ContextCode by ContextStream

A coding agent that remembers your project.

Most AI coding tools start every conversation from zero. ContextCode starts from your project's durable memory: the decisions you made, the sessions you already had, the code graph, and the files that actually matter for the task in front of you. Every turn shows you what it was grounded in, and every change it makes is reviewable, checkpointed, and reversible.

ContextCode runs the agent locally through a bundled engine and grounds it with ContextStream.

What you get

Grounded answers, with receipts. ContextCode retrieves relevant context automatically and shows a source-aware receipt for each turn: what it used, how fresh it was, and where it came from. You can pin, unpin, attach, and detach sources directly from the receipt.

Memory that outlives the session. Decisions, notes, and insights you approve are saved to ContextStream and resurface in later sessions, in other tools, and for teammates on a shared workspace.

Governed autonomy. Choose a posture per run: read-only analysis, ask before every change, safe autonomy for routine workspace work, or an explicitly confirmed bypass. Shell, network, package, external-service, credential, destructive, and publication actions always ask or stay blocked, regardless of posture.

Review before you keep anything. Structured edits are checkpointed and grouped into change sets you inspect in native VS Code diffs against exact baselines. Restore a checkpoint if you don't like where a run went.

Parallel agents in isolated worktrees. Launch Build, Explore, and Review agents alongside your main session. Write-capable agents run in their own Git worktrees so they can't collide with your working tree, and you review before merging.

Ask, Plan, Agent, and Autopilot modes plus Auto, Coder, Planner, Designer, and Prototyper prompt roles, a keyboard-accessible slash palette, image attachments, streaming responses, cancellation, and crash recovery from a local journal.

Requirements

• VS Code 1.95 or later on macOS, Linux, or Windows (x64 or arm64).
• A ContextStream plan that includes ContextCode. Agent access starts at Pro. See contextstream.io/pricing.
• A trusted, file-backed workspace. ContextCode renders its interface in an untrusted workspace but will not launch the engine, read workspace files, or send prompts until you trust the workspace. Virtual workspaces are not supported yet.

The matching ContextCode engine ships inside the platform-specific package, so there is nothing else to install. The extension verifies the engine's target, protocol version, and SHA-256 checksum before launching it.

Getting started

1. Install the extension and open a workspace folder.
2. Trust the workspace when VS Code asks.
3. Open ContextCode from the Activity Bar.
4. Run ContextCode: Sign in to ContextStream from the Command Palette. The current engine uses ContextStream's CLI device-flow sign-in, so ContextCode offers to open a secure sign-in terminal for you.
5. Ask for something. ContextCode grounds the request, plans, and shows you the work as it goes.

You can also talk to it from VS Code's chat view with @contextcode, using /ask, /plan, /agent, or /review.

The sidebar

Settings

Language model tools

ContextCode contributes four read-mostly tools you can reference from VS Code chat: #contextStreamReceipt for grounding provenance, #contextStreamCodeSearch for the project code index, #contextStreamCodeGraph for code-health inspection, and #contextStreamMemory to save a durable record after an explicit confirmation.

Privacy and safety

• Grounding is scoped to the workspace folder you selected.
• Credentials live in VS Code's secret storage and are never written to prompts, journals, evidence bundles, or webview state.
• Image bytes are carried as opaque handles and never land in prompts, journals, evidence, or snapshots.
• Repository-committed policy files may add ask and deny rules but can never grant the agent authority. Committed configuration is not treated as your consent.

Current limitations

We would rather tell you these up front:

• Shell commands always require targeted approval, including tests and builds. A repository-controlled script is arbitrary code, so automatic shell execution stays off until the engine has a verified OS-level sandbox. Shell, browser, network, and external-service approvals are once-only, because the same command can execute different code after the workspace changes.
• Checkpoint restore covers ContextCode's structured file tools. An approved shell command can make changes outside that capture path.
• Durable allow rules are disabled until they can be stored in a user-owned, workspace-bound policy store.
• In-editor plan/start and agent/steer are still on the roadmap. Sign-in uses the CLI device flow shown above.
• A hosted/BYOK language model chat provider and managed cloud agents are later milestones.
• The full Stable/Insiders remote and OS matrix (SSH, Dev Containers, Codespaces), the accessibility audit, and the performance and reliability gates are still in progress.
• Across an engine restart, side-effecting bridge calls are at-least-once unless the underlying operation enforces its own uniqueness.

Support

• Questions, bugs, and feedback: contextstream.io/contact
• Product and docs: contextstream.io

License

ContextCode is proprietary software. See the LICENSE file included in the package. Use requires a valid ContextStream subscription or a separate written agreement.