Switching between Claude Code, Codex, and Antigravity means explaining your progress from scratch every time you change models. AgentBridge opens the three CLIs as tabs in one workspace on a shared record of the work: a running summary, the raw turns behind it, and long-term memory about you and the repository.
That record is never pushed into the prompt. The hook adds a single short instruction, and the agent pulls what it needs through the agentbridge command when it decides it needs it.
Who it is for
People who repeat the same explanation while moving between the three CLIs
People who want several agents on the same task in one window
People who want to use the CLIs and subscriptions they already have, with no new account or server
Features
One workspace, three agents
Claude, Codex, and Antigravity tabs run side by side. The session tree shows which sessions are working and which have finished.
Context that survives a model switch
The summary, the raw turns, and what you have told the agents live in one record, and a new tab pulls only the part it needs — so you do not restate the situation. The injected block runs about 1KB, small enough that recent turns are not pushed out.
Agents that launch agents
Ask the agent you are talking to for a helper. The helper opens in its own tab, nested under the parent session, and the parent sends follow-up instructions and reads what the helper recorded.
Isolated worktrees
A helper can run in its own git worktree, so two agents editing the same file never collide. While it runs, the worktree appears in the built-in source control view; when it finishes, the changes land in one go — and if a single conflict shows up, nothing lands and your folder is left as it was.
Memory that needs your approval
Only durable things are kept — your role, how you work, the conventions of the repository. The agent proposes as it works, and nothing becomes memory until you approve it.
HTML viewer
View HTML pages created by agents inside the IDE without opening a browser. Click an element on the page to attach a note and send it to a session. The view updates automatically whenever the file changes.
Also included
Your CLIs, your subscriptions — no server of ours, no separate account. Model costs stay inside the CLI subscriptions you already pay for.
Subscription quota display — remaining subscription quota for Claude, Codex, and Antigravity appears as icons under the Sessions header. Hover for a tooltip with the weekly remaining quota and reset time.
Your project folder untouched — hooks and skills install into your own agent settings. No AgentBridge file is written to the repository.
Cheap background work — summaries and session naming run headless on a CLI you pick. When its quota runs out, the next CLI takes over.
Drag paths in — hold Shift and drop a file on the chat, and its path lands in the input line as @path. Explorer files, editor tabs, and files from outside the IDE all work.
Sessions that reopen — reopen the IDE and each CLI picks up where it left off through its own --resume.
CLI behavior preserved — permission prompts, tool approval, and session management stay the way each CLI does them.
Supported agents
AgentBridge runs the CLIs installed on your machine. At least one of the three is required.
To authenticate, run each CLI once and follow its prompts. Antigravity uses agy /auth.
To keep background work on a free quota, install and authenticate Antigravity (agy). Without it, background work falls back to the model you are talking to and warns you about the token use.
Install
Search for 'AgentBridge' in the extensions tab. Cursor, Antigravity IDE, and Windsurf use their own extensions tab the same way.
Once installed, an AgentBridge icon appears in the activity bar. Clicking it opens three panels in the sidebar: Sessions, Context, and Long-term Memory.
Press the + button in the Sessions panel title bar. The AgentBridge icon at the top right of an editor tab does the same thing.
Pick a model from the list, and a chat tab opens with that CLI already running.
Sessions collect in the Sessions panel. Click one to reopen it, or hover a row to rename or delete it.
The Context panel holds the running summary and the raw turns; the Long-term Memory panel holds proposals waiting for approval.
When you need a helper, just say so to the agent in the chat tab. Launching one, checking on it, reading its result, and merging its changes are already in its skills.
How it works
Your local CLIs, run directly — AgentBridge starts the CLIs you have already authenticated in a terminal session of its own. No relay server sits in between.
Context on demand — the hook carries a short instruction rather than a copy of your memory. The agent pulls the summary or the recent turns it needs, so the prompt stays small and the record stays whole.
Your project folder intact — hooks and skills install into a marked block inside your own settings, leaving the rest alone. Everything is stored under ~/agentbridge/, and agentbridge uninstall removes the hooks.
Shift-drag to insert paths — the chat intercepts a drag ahead of the IDE only while Shift is held. Paths from the explorer and editor tabs are read as they are; files from outside the IDE are copied into ~/agentbridge/attachments/ and that path is inserted. The notation is @path, which all three CLIs understand.
Settings
VS Code settings (settings.json or the settings UI):
Key
Default
Description
agentbridge.refine.policy
active
Policy for picking the CLI that handles background work
agentbridge.refine.priorityOrder
[agy, codex, claude]
CLI try order under the priority policy
agentbridge.refine.fixedCli
agy
CLI pinned under the fixed policy
agentbridge.turns.assistantDetail
compact
How much of each answer to keep in the turn record
agentbridge.memory.maxArchiveSnapshots
15
How many previous snapshots to keep
Where data lives
Everything is stored under ~/agentbridge/, split per project folder.
~/agentbridge/
├── workspaces/<folder>-<hash>/
│ ├── workspace.json ← sessions and state
│ ├── ir.json ← the running summary
│ ├── turns.jsonl ← raw turns
│ ├── archive/ ← previous snapshots
│ ├── sessions/<id>/ ← per-tab replay buffer, hook signals
│ └── trees/<name>/ ← isolated subagent worktrees
├── attachments/ ← files dropped into the chat
└── global/
├── profiles/default/ ← what is known about you
└── projects/<repo>/ ← what is known about this repository
Each agent's global settings file gains one marked block for the hooks and skills, and nothing else changes.
AgentBridge runs no server of its own. Your messages go through the CLIs you authenticated, to the backends those CLIs already talk to (Anthropic · OpenAI · Google). Background work such as summaries and session naming runs on the same CLIs, and the results are stored on your machine.
Nothing else leaves your machine. There is no analytics, no telemetry, no third-party service. Session records, turn logs, memory, and terminal replay buffers are all local files.
