The agent in your editor:
sees what you see, changes what you approve.
The extension runs a local MCP server inside your editor; cc agent in the integrated terminal connects automatically —
editor capabilities become agent tools, and agent edits go through the editor's native diff under your judgment.
VS Code 0.37.109 is public on Open VSX with 46K+ downloads and recommends npm latest CLI 0.166.65; JetBrains 0.4.130 is public on Marketplace and recommends the same CLI. Public listings and install artifacts were read back. No Microsoft VS Code Marketplace build is published.
Public VS Code 0.37.109 / CLI 0.166.65 automatically save bounded task notes. `↗` / Continue in new chat waits for a completed tool boundary before opening an independent session. It copies neither the full prior transcript nor temporary permissions, and invalid history fails before a model call.
Inside your editor.
Durable task notes and a fresh session
Open VSX 0.37.109 and JetBrains 0.4.130 with CLI 0.166.65 save goals, file evidence, failures, and next steps as authoritative checkpoints plus a bounded WORKLOG.md; handoff copies neither the old transcript nor temporary permissions.
Bounded reads and loop recovery
Open VSX 0.37.109 with CLI 0.166.65 admits a bounded read of a known target during recovery. Synthetic pauses no longer feed repeated-read or identical-output counters, while genuine repetition and unrelated remote discovery still fail closed.
Selection with every prompt
Each prompt automatically carries your current selection + open tabs (an <ide-context> block) — the model knows what you are looking at, never persisted to the session.
Diagnostics fed back
After the agent edits a file, language-server errors/warnings flow back into the tool result — it sees and fixes its own breakage within the same loop.
Native diff approval
File-edit confirmation upgrades to the editor’s side-by-side diff: edit the right side, Accept replaces execution, hunk-level "Pick hunks…" partial accept, "Request changes…" inline notes that the agent revises against; a cross-file change can go through openMultiDiff for one batched multi-file review. Reject never writes.
Jupyter kernel
executeCode runs code in the active notebook’s kernel and returns output, variables persist across calls (VS Code).
Sidebar Chat panel
Talk to the agent without a terminal: streaming replies + tool traces + Plan-mode cards + Approve/Deny cards + multiple conversation tabs (each its own agent process, switching restores the transcript) + session resume + /sessions switcher.
Exact scoped approvals
Open VSX 0.37.109 keeps one-shot approval as the fast path and adds explicit current-turn, current-session, deny, and cancel choices. Turn/session grants reuse only the exact capability, scope, and binding from the CLI request; the Webview cannot widen authority.
In-editor app preview
Start App Preview finds your dev script → runs the dev server → detects the served URL → opens it embedded in the editor (VS Code’s built-in Simple Browser / JetBrains’ JCEF browser, falling back to the external browser when JCEF is unavailable); the dev server’s HMR handles live reload on edits; if the dev server dies unexpectedly it reports a crash with a one-click Restart instead of silently stopping.
Terminal-context sharing
getTerminalOutput lets the agent see the commands you just ran in the integrated terminal — their output and exit code (VS Code 1.93+ shell integration) — so it knows what you ran and how it failed. Auto-injected into <ide-context> each turn, plus an explicit @terminal mention.
Tests, coverage, and debugger context
cc-ide-quality/v1 sends bounded quality snapshots with freshness metadata. Stale results stay visibly stale, and VS Code notebook execution carries the real notebook and cell context.
Transactional upgrades and supply-chain governance
Enable or disable by scope. Upgrades stage and revalidate manifests/signatures/SBOMs before atomic activation, restoring the previous version on failure. IDEs reload only after activation and require explicit consent for capability widening.
Distributed Agent Teams
The IDE observes raw local team-state v6 and separate distributed-queue v1 snapshots read-only. Takeover, managed-checkpoint recovery, and side-effect adjudication return to the CLI with authority-digest, lease, and evidence fences. Stale UI actions fail closed; the IDE never mutates authoritative JSON directly.
Generated events, content-free projection
Open VSX 0.37.109 and JetBrains 0.4.130 use Schema-generated payload unions while preserving unknown future events; VS Code retains bounded canonical message/follow-up and custody-handoff counters.
Fixed-capability App Server pilot
VS Code can explicitly enable the default-off App Server pilot. It exposes fixed Thread/Turn methods rather than arbitrary JSON-RPC; approvals decline until the review UI lands, and failure never bypasses CLI permissions or sandboxing.
Tool duration and retries
Usage views show measured tool duration, observed same-turn retries, and secret-free streaming LLM retry reasons with actual provider/model identity, without retaining tool arguments, output, or credentials.
Governed Automation Center
Open VSX 0.37.109 and JetBrains 0.4.130 show CLI-owned flow/Routine projections, scope, preflight, and history. Mutations use exact revisions and CLI argv; the IDE never writes authoritative storage directly.
Workbench, PM readiness, and signed deployment
Open VSX 0.37.109 and JetBrains 0.4.130 verify, persist, and pause deployment profiles through the CLI, then show budget/signer, SQLite-seal, transition-recovery, and snapshot-store configuration read-only. IDEs receive no runner, grader, database bytes, Ledger, or governance private keys; PM readiness remains non-executable and automatic active promotion remains on HOLD.
Paginated read-only Skill Library
Filter bundled, marketplace, managed, and workspace Skills by source and state, with canonical digests, index witnesses, and verified-outcome ranking. Browsing never installs or promotes a Skill.
Audited Artifact workbench
Preview, open, copy-path, reveal, and download first pass CLI artifacts access, which revalidates managed bytes and writes a content-free audit event. Deletion uses a stable id so response-loss recovery does not issue a second destructive operation.
CLI-owned Sessions Workbench
Public 0.37.109 / 0.4.130 consume immutable CLI session/message projections; CLI 0.166.65 provides a paged read-only index with an authoritative full-scan fallback for large background histories. Stale actions still fail closed.
Resumable delivery and rewind
The public build covers GitHub, Gitee, configured remotes, and manual handoff. Each step verifies result/effect digests; rewind binds session, workspace, repository head, checkpoint revision, and manifest digest.
Request-level Context Center
Public 0.37.109 / 0.4.130 surface execution location, workspace authority, session, workflow, and recovery identity. With CLI 0.166.65 they use the actual primary-request window and input/output/cache usage, exclude attributed child calls, and label fallback estimates.
Side-effect Center
Both IDEs summarize permission activity plus scoped grant and revoke entry points. Every change still executes through CLI commands and exact context; the IDE never bypasses authoritative storage.
Ask and edit beside a selection
Open VSX 0.37.109 provides isolated inline sessions, streaming replies, code-block actions, and interaction recovery while retaining the stable/minimum × Linux/Windows/macOS ARM64 multi-window isolation gate. Explain / Refactor / Fix / Docs / Tests are public.
CLI version pairing
npm latest is chainlesschain@0.166.65 at commit 17509017a2. Its three-OS CLI CI, Strict Sandbox, OIDC provenance, and public fresh-install readback passed. Open VSX 0.37.109 and JetBrains 0.4.130 both recommend this CLI; Microsoft Marketplace is not published.
Governed automatic ghost text
Manual Alt+\ completion is unchanged. The default-off automatic path adds cancellable 650 ms debounce, exact-context dedupe/cache, separate hourly request and context budgets, bounded output, and quality fallback. Results slower than five seconds or below quality fail quiet.
The extension is just the bridge —
cc agent does the work.
Run cc agent in the editor's integrated terminal and it auto-connects to the extension's MCP server — that's how the capabilities above come alive.
The same agent ships a full suite of companion commands: a permission gateway, checkpoint rollback, cost reporting, loop automation and structured output — identical inside the IDE and in a bare terminal.
cc agent Auto-connects to the editor inside the integrated terminal — selection injection, diagnostics feedback, native-diff approval all live
cc agent -p "…" Headless single turn for CI / task scripts (text / json / stream-json output)
cc ide doctor Diagnose why the bridge did / did not connect (no lock / stale / workspace mismatch)
cc permissions .claude/settings.json allow / ask / deny rules + credential-read guard
cc checkpoint Auto-snapshot inside a git repo; one-click restore when an edit goes wrong
cc cost / cc loop / cc goal Spend reporting, scheduled loops, persistent cross-session goals
Two steps to install
- 1. Install the CLI:
npm i -g chainlesschain@0.166.65(current npm latest0.166.65, bridge minimum0.162.47;cc ide doctordiagnoses connectivity) - 2. Install the extension: VSCodium / Cursor etc. — search ChainlessChain IDE in the extension panel (Open VSX currently
0.37.109); IDEA / PyCharm / WebStorm — Settings → Plugins → Marketplace, same name (current0.4.130, 2024.2+). Stock VS Code does not query Open VSX, so do not use Open VSX's generic Install link. Download the versioned 0.37.109 VSIX, then run Extensions: Install from VSIX....
Safe defaults: server binds localhost only + per-instance random Bearer token + 0600 lockfile; on the CLI side --no-ide / CC_IDE_CONTEXT=0 / CC_IDE_DIFF_APPROVAL=0 switch each behavior off.
Protocol, troubleshooting, architecture
Discovery protocol (lockfile + env), Streamable HTTP, editor-agnostic design — all in the docs.
Read the IDE bridge docs ↗