AI Context Bridge

English · ภาษาไทย

Shared blackboard memory + skill control center for multi-model AI workflows. Runs in VS Code and any VS Code-based editor: Cursor, Windsurf, Google Antigravity, VSCodium, Trae, Void, code-server, etc.

When you switch from Claude → GPT → Gemini mid-task, you usually have to re-explain everything. AI Context Bridge keeps a small, local JSON blackboard that any agent can read and write — so the next model picks up where the last one left off.

It's pure state management. No API calls. No token math. No cloud sync. Just a transparent, local memory you control.

What you can do with it

• Hand off work between AIs without re-explaining — Claude finishes, then GPT picks up exactly where it left off (same pinned files, same recent thoughts, same skill rules).
• See what every model thought, in order — a single Thought Timeline with model badges and handoff hints.
• Pin the files that matter right now — 📌 badges on tabs and Explorer; auto-pin recent edits; spec files (CLAUDE.md, AGENTS.md, .cursorrules, …) are pinned automatically.
• Decide which AI capabilities are allowed — every "skill" has a 3-state toggle: ✅ Enabled / ❓ Ask / ⛔ Disabled.
• Sync targets without hunting through views — Quick Actions includes one Target Settings picker for skill mirrors, context bridge files, and MCP copy targets.
• Repair generated context blocks — Sync All Now deduplicates AICB-managed blocks while leaving user-written text outside the markers alone.
• Roll back if an agent breaks something — one-click snapshots of the entire memory state.
• Stay private — everything is local JSON in your workspace. No telemetry, no cloud, no account.

Install

VS Code Marketplace

Search AI Context Bridge in the Extensions panel, or:

code --install-extension jeurboy.ai-context-bridge

Open VSX (Cursor / Windsurf / VSCodium / code-server / Theia)

Search AI Context Bridge in the Extensions panel — most non-Microsoft editors use Open VSX as their default registry.

From .vsix (any VS Code-based editor)

Download the latest .vsix from Releases, then:

# VS Code / VS Code Insiders
code --install-extension ai-context-bridge-0.6.2.vsix

# Cursor
cursor --install-extension ai-context-bridge-0.6.2.vsix

# Windsurf
windsurf --install-extension ai-context-bridge-0.6.2.vsix

# VSCodium
codium --install-extension ai-context-bridge-0.6.2.vsix

Or via UI on any VS Code-based editor (including Google Antigravity, Trae, Void): Extensions panel → ... menu → Install from VSIX...

Editor compatibility

Uses only the stable VS Code Extension API (vscode ^1.85.0), so it works in any VS Code-compatible host:

First 5 minutes

1. Open the sidebar. Click the 🧠 AI Context Bridge icon in the Activity Bar. You'll see five views: Quick Actions, Skills, Pinned Files, Snapshots, MCP Servers.
2. Let auto-discovery do its thing. Skills under .claude/skills/, .cursor/, .gemini/, .codex/, .agent/ (and their ~/... global counterparts when enabled) are detected automatically. Spec files (CLAUDE.md, AGENTS.md, .cursorrules, README, ARCHITECTURE.md, plans/, rfcs/, …) are pinned to the Spec / context group on activation, including matching files inside nested subprojects. MCP servers are pulled from each host's config file (.mcp.json, ~/.claude.json, Claude Desktop, Cursor, Gemini, Windsurf, VS Code, Kilocode, Codex).
3. Hit Sync All Now. The big button in the Quick Actions panel writes the AICB block into every configured agent file (CLAUDE.md, AGENTS.md, .cursorrules, …), mirrors skills cross-agent, and refreshes the MCP inventory in one click. The same button also lives on the Pinned Files and Snapshots toolbars, plus the status bar.
4. Hand off when needed. Use Copy Bootstrap Prompt in Quick Actions to copy a paths-only prompt into agents that read files themselves (Codex CLI, Aider). When you're switching agents mid-session and the new one already worked here, use Copy Reload Prompt instead — it tells the agent to refresh from .aicb/state.json and re-read recent thoughts so it doesn't act on stale context. For agents in a chat box that don't have file access, the Skill API also exposes copyHandoffPrompt programmatically.
5. (Optional) Tune the targets. Use Target Settings in Quick Actions to choose skill mirror targets, context bridge files, and default MCP copy targets in one picker.

This extension is sync-only — it does not manage skill state. Skill statuses (ENABLED / ASK / DISABLED) are determined automatically by name/description (skills with exec/run/delete/push/etc. become ASK). The UI displays them; you cannot edit them. To exclude a skill, remove it at the source (delete from .claude/skills/...).

Daily usage

The sidebar

Open the 🧠 icon in the Activity Bar.

Skills view — every AI capability auto-discovered for this workspace (read-only).

• Toolbar: 🔁 Rescan skills · "..." overflow (timeline, mirror, configure mirror targets)
• Status (✅ Enabled / ❓ Ask / ⛔ Disabled) is determined automatically from each skill's name/description; the UI displays it but does not let you edit it. Risky-looking skills (anything matching exec | run | delete | push | …) default to Ask.
• A skill set to Ask triggers a HITL modal when an agent calls the public hitl.authorize() API: Allow once / Allow this session / Deny

Pinned Files view — files in working memory, split into:

• Spec / context — auto-imported, never expire (CLAUDE.md, AGENTS.md, .cursorrules, README, plans/, …)
• Working memory — recent edits, dwell-pins, and files you pinned manually
• Toolbar: 📌 Pin current file · 📋 Copy handoff prompt · 🔄 Rescan specs · ↔️ Bridge to agent files now
• Right-click → Unpin to remove

Snapshots view — points-in-time of the whole memory state.

• Toolbar: 💾 Create snapshot
• Click ↩️ to restore, 🗑️ to delete

The status bar (bottom-left)

• 🧠 N skills · M pinned · K thoughts — quick stats; click to Force Sync to disk
• 🕒 Timeline — click to open the Thought Timeline

Pinning a file

• From the editor: click the 📌 icon in the editor title bar
• From a tab: look for the 📌 badge — present means pinned
• Auto: files you save (or that you keep open for autoPinDwellMinutes) get auto-pinned and expire after autoPinExpireMinutes of inactivity. Manual pins never expire.

Handing off to another AI

The fastest way: click the Copy Handoff Prompt button on the Pinned Files view title bar → paste into the next chat.

The "always-on" way: enable aiContextBridge.exportToAgentFiles. The extension writes a managed <!-- AICB:BEGIN ... AICB:END --> block inside whichever convention files exist (CLAUDE.md, AGENTS.md, .cursorrules, .windsurfrules, GEMINI.md, .github/copilot-instructions.md). Edit the rest of those files freely — only the marked block is touched.

By default, files outside the marked block are never created. Set agentFilesOnlyExisting: false if you want missing convention files auto-created.

Per-editor convention files

When exportToAgentFiles is on, this is what each AI reads automatically:

• Claude Code / Claude Desktop → CLAUDE.md
• Cursor → .cursorrules, .cursor/rules/**/*.{md,mdc}
• Windsurf → .windsurfrules
• GitHub Copilot → .github/copilot-instructions.md
• OpenAI Codex → AGENTS.md at repo root (project guidance), ~/.codex/AGENTS.md (global guidance — opt-in via aiContextBridge.agentFiles)
• Agents SDK / legacy → AGENT.md
• Google Gemini / Antigravity → GEMINI.md
• Kilocode → .kilocoderules, .kilocode/rules/aicb.md
• Generic agent (.agent) → .agent/AGENTS.md, .agent/rules/aicb.md

Tweak the list with the aiContextBridge.agentFiles and aiContextBridge.specPatterns settings.

Agent path reference

The full map of where AI Context Bridge reads from (discovery) and writes to (mirror / sync) for each host. Workspace paths are relative to the project root; global paths are under the user's home directory.

Skill discovery + mirror

Mirror targets are gated by the aiContextBridge.mirrorSkillsToOtherAgents setting. Generated files carry an AICB:GENERATED marker — they're auto-pruned when the source disappears, and AICB will never overwrite a hand-authored file at the same path.

MCP server discovery + sync

Per-server copy uses these exact files as targets. Synced entries are tagged with _aicbGenerated: true + _aicbSource: "<host>:<scope>" so they're identifiable on the next round-trip. Hand-authored entries with the same name prompt before overwrite.

Note (Codex CLI): the official Codex CLI stores MCP servers in ~/.codex/config.toml (TOML format). AI Context Bridge currently only reads/writes JSON, so Codex sync uses .codex/mcp.json as a parallel JSON file rather than the canonical TOML. Native TOML support is on the roadmap.

Handoff context (.aicb block)

The bridge only updates content between the <!-- AICB:BEGIN --> and <!-- AICB:END --> markers — anything else in those files is preserved verbatim. Each generated block carries checksum metadata. Sync All Now repairs duplicate managed blocks by collapsing them back to one fresh block; background auto-sync preserves checksum-mismatched blocks so user edits inside the managed area are not silently overwritten. Manual duplicate text outside the markers is left untouched for user review, not deleted automatically. By default, only files that already exist are updated (agentFilesOnlyExisting: true).

Where AI Context Bridge stores its own state

Override the location with the aiContextBridge.storagePath setting.

UI controls (no Command Palette needed)

Every action is reachable from a button — the extension is intentionally not exposed in the Command Palette. The 5 views in the activity bar give you everything:

Settings

Tips

• Don't see your skills? Click 🔄 Rescan Skills on the Skills view, or check that .claude/skills/<name>/SKILL.md files exist.
• Status bar shows wrong counts? Click the 🧠 status item to Force Sync.
• Switching machines? state.json is plain JSON in <workspace>/.aicb/. Commit it (or don't) — it's yours.
• Exporting context to a fresh chat? Use Copy Handoff Prompt. It's the single fastest way to bring a new AI up to speed.

Why this exists

Modern coding flows mix multiple models. Each one has its own scratch context, and human-in-the-loop boundaries get lost between handoffs. AI Context Bridge gives you:

• Visibility — see what each model thought, in order, with model badges
• Control — gate any tool call behind a modal, or freeze it entirely
• Continuity — the next agent reads state.json and just keeps going

Everything is local. There is no telemetry, no network sync, no cloud account. Privacy by design.

For extension authors

Other extensions can drive the blackboard via a small public API. See docs/API.md for details, or:

const ext = vscode.extensions.getExtension('jeurboy.ai-context-bridge');
const api = await ext?.activate();
api.memory.addThought({ modelId: 'claude-opus-4-7', text: '…' });

You can also write <workspace>/.aicb/state.json directly — every change is picked up automatically.

License

MIT — see LICENSE.