Chat with Claude Code and Codex, review saved changes, manage accounts and track tokens — inside VS Code.
Keep working in your native agent or use the built-in chat. AgentTandem brings saved file changes, conversation history, account switching and usage into one workspace companion.
| Feature | What you can do |
|---|---|
| Code and image review | Compare saved snapshots, accept or reject individual blocks, navigate files and undo the latest revert. |
| Chat and history | Run Codex or Claude Code, resume project conversations, attach files and queue follow-ups. |
| Account management | Save and switch Codex, Claude Code, Cursor and Antigravity sign-ins. |
| Token dashboard | Explore account, model and conversation totals, cache usage and estimated API costs. |
| Live quota | See remaining allowance, reset times, observed burn rate and quota estimates. |
| Checkpoints and storage | Preview code restores, see what agent history takes up, delete old Codex/Claude conversations and track how much space you have freed. |
- Install AgentTandem in VS Code and open your project. For chat, install and sign in to the Codex or Claude Code CLI you want to use.
- Select AgentTandem in the Activity Bar. Check Tracking on in the status bar before asking an agent to edit. If tracking is off, click it, choose Resume tracking and wait until it is ready.
- Work in your usual agent tool, or select + to start a built-in chat. Manage CLIs lets you choose an installed CLI.
- Save your changes, open Review and select a file. Accept keeps the code and marks it reviewed; Reject / Revert restores the saved baseline. Accept does not create a Git commit.
Tracking captures saved, non-ignored files, including manual edits. It cannot reconstruct changes made before tracking starts. Running agents requires a trusted workspace and an authenticated CLI.
Review opens with changed files grouped by conversation. Shared and unassigned changes stay visible when their source cannot be established. Select a file to compare immutable before/after snapshots in the bundled Monaco editor, with syntax highlighting, search and inline actions.
- Accept / Reject acts on an individual change block. Accept all keeps all changes in the current file; Reject all restores the whole file to its saved baseline after a confirmation that names the file. Newer saved edits and unsaved buffers still block rejection, and Undo recovers it.
- Previous / Next change and Previous / Next file move through the review. Use the file picker to jump directly to another file.
- File and conversation Revert actions preview their scope. Undo recovers the latest file, group or block revert.
- … contains Refresh, Native diff and conversation details. Images show saved Before / After versions with whole-file actions.
- Image changes opened from Source Control (and Git image diff tabs restored after a reload) open in AgentTandem's Before / After view instead of VS Code's built-in Git image diff, which in VS Code 1.142 fails on its original side (
Cannot read properties of undefined (reading 'selection')) and closes. Staged changes compare HEAD → index, unstaged index → working tree, with Previous / Next file across the changed images. AgentTandem: Open Image Changes (Source Control context menu, image editor title, Command Palette) opens it directly; setagentTandem.openGitImageDiffstofalseto keep the built-in diff.
Run AgentTandem: Manage Accounts, or open Settings → Accounts. Filter by app and search saved profiles for Codex, Claude Code, Cursor and Antigravity.
- Select Add account and choose an app. For Codex or Claude Code, sign in through the native CLI, Save current sign-in, or Import JSON. For Cursor or Antigravity, open the installed app, sign in there and save its current sign-in.
- Select Switch account to update the default CLI or IDE sign-in. Finish active agent turns first; close Cursor/Antigravity windows before switching and reopen the app afterward.
- Follow the progress dialog and timestamped log. Hide keeps the operation running; View switch log reopens the latest result. The switch saves the outgoing sign-in, verifies the selected identity and attempts rollback if verification fails.
After a Codex or Claude Code switch, select Restart Extension Host to let the native VS Code extension pick up the new sign-in, or Later to defer it.
Hover a provider's quota in the VS Code status bar to open the account menu: remaining quota, reset countdowns, saved accounts and Open Dashboard are available without clicking the status bar. Select a saved account to switch. Refresh quota stays in the menu; clicking the status bar opens a searchable account picker with the same labeled shortcuts. In the dashboard, View usage selects an account's statistics without switching its sign-in.
Quota rows use theme colors: green for healthy, amber at 25% remaining or less, red when exhausted. The active account, reset actions and extension storage have distinct icons.
The hover and status bar use the same current quota snapshot. Repeated limit windows and saved copies of the same account appear once; hover percentages show quota remaining, and saved quota is marked Cached quota. A profile named after its email does not repeat that email on the next line.
Saved-account quota is checked in the background at most once every ten minutes, even with Settings closed. Opening the status-bar hover checks the same cache when VS Code enables statusBarItemTooltip; other builds use the background snapshot. Manual refresh has a one-minute cooldown. Reads are sequential, overlapping refreshes share one job, and failures back off from ten minutes up to two hours with jitter. A longer Retry-After is honored when the CLI exposes it, following OpenAI's error guidance and Anthropic's rate-limit guidance. These cache intervals are AgentTandem defaults, not provider-published polling limits.
Saved profile identity is checked against credential metadata before native sign-in is mirrored. A different account cannot overwrite a saved profile or supply its plan/quota. Expired sign-ins keep their saved identity and require signing in again.
For the active Codex ChatGPT account, Reset usage uses one earned reset to restore eligible quota. The available count is shown beside the action. After redemption, AgentTandem reads quota from Codex again and reports the result. The action is disabled when no earned resets are available or an agent turn is active; token history stays intact. The same action is available in the account picker and dashboard account cards. Requires a selected Codex CLI that supports earned resets. See OpenAI Docs: earned rate-limit resets.
Saved credentials stay in local native-auth profile storage and are not sent to the webview or included in switch logs. Removing a saved profile keeps the default native sign-in and conversation history. Cursor/Antigravity quota stays in the native app; Antigravity 2 OS-vault switching currently supports Windows.
Account switch progress
Run AgentTandem: Open Dashboard, or use Open Dashboard in the sidebar toolbar. The dashboard opens on Today for all accounts.
- Filter bar — date chips (Today to All time, Custom, 5h cycle, Weekly cycle), then provider, account, model and grouping by day, week or month.
- Summary — tokens, API-equivalent cost, quota and Unknown account tokens for the current selection, with Assign when some history has no account.
- Tabs — Overview (activity chart and account cards), Quota (current quota, burn-down and tokens per 1% quota), Conversations, and Details (API activity, token totals and cost, period and account tables, history assignment and data notes).
View usage on an account card filters the whole dashboard to that account. Select ← All accounts to return to the cards.
5h cycle and Weekly cycle show one account's usage between its quota resets, using the chosen account or the active sign-in. A cycle runs for its window length up to the reset; an early reset (a banked or purchased reset) ends it where the next one starts. Use ‹ › to step through recorded cycles and Current cycle to return. The chart shows each hour of a 5h cycle or each day of a weekly one, and Unknown account shows that provider's unattributed tokens in the same window, so you can see what the account total may be missing.
Cycles are known only from quota AgentTandem has read, so cycles from before it first checked an account are not listed.
| Source | What it shows |
|---|---|
| API | Native account quota and, when supported, Codex account token activity: lifetime tokens, peak day, turn duration, streaks and returned daily totals. |
| Local | Deduplicated token counts recorded in local Codex and Claude history across projects, including archived and subagent sessions. |
| Estimate | API-equivalent USD cost, tokens per 1% quota and projected reset-cycle capacity. These are estimates, not subscription charges or invoices. |
Details shows API token activity (lifetime tokens, peak day, longest turn and streaks), Local input/output/cache totals, the API-equivalent cost split, the period breakdown by model and usage by account.
API activity and Local totals are displayed separately and never added together. Codex API activity depends on CLI/account support; Claude token details come from local history. Model filters apply to Local usage. Missing API metrics remain unavailable, and unknown model prices remain Unpriced.
Account attribution preserves native per-response and per-turn evidence across account switches within a conversation. Codex responses without that evidence are Matched from quota changes when the quota reset times Codex recorded with them match exactly one account's saved quota readings. Claude responses use the session owner that recent Claude Code versions (for example 2.1.319) record in the transcript, labelled Inferred from session evidence. Other responses are matched when exactly one account's quota rose around them; this is an inference from AgentTandem's quota samples, so history from before those samples is not matched. Responses still without an account use the Active CLI sign-in AgentTandem read around that time (about once a minute while a window is open); this assumes the same machine's CLI and editor extensions used that sign-in, so a different account used in another terminal or config directory is attributed wrongly. Time just after a switch is left unknown until the extension host restarts, and so is time no window was watching. Before that fallback, a Codex week AgentTandem never read quota for is Inferred from session evidence when every thread started in it was created by the same known account (a thread's first response is billed to its creator; later turns may not be). Other history stays Unknown account. Assign history to an account lets you confirm a conversation/date range or keep it unknown without rewriting native transcripts, and takes priority over a match.
On the Quota tab, each quota window (Codex/Claude Code × 5h/7d) gets a card plotting remaining quota in the current reset cycle against an even pace to reset, plus a projection from the recent burn rate. The card says whether the window is On track (and how much will be left at reset) or Runs out ≈ time · before reset, with even-pace, per-cycle and last-hour rates.
Each provider has separate 5h and Weekly windows. Every window starts with four figures side by side: the tokens per 1% rate (with its source, cycle count and change against earlier cycles), the estimated full cycle, the tokens left in this cycle and the current cycle's quota interval. The full cycle is approximately how many tokens a full reset cycle represents: recorded tokens ÷ quota percentage points used × 100. It uses the current cycle's recorded ratio when available, or the weighted average of recorded cycles, with the basis shown below the estimate.
The interval figure shows the quota readings and their recorded tokens, such as 100% → 75% remaining: 804,531 tokens. Recording that starts partway through a cycle shows that actual interval, such as 84% → 82%. Older API records without a saved starting reading show Start unknown.
Use Chart metric to switch between:
- Tokens / 1% — a line chart of cycle ratios, with the weighted average as a dashed line.
- Recorded tokens — bars showing the tokens actually recorded in each cycle.
- Full quota estimate — bars extrapolating each observed ratio to a full cycle.
The cycles are part of the chart: the figures for each one sit in a table directly under its column, so there is no separate list. A column shows the cycle start (marked Current, Partial or Legacy where that applies), quota remaining, recorded tokens, quota used in percentage points, tokens / 1%, the full-quota estimate and the change against the previous cycle. The row for the chosen metric is emphasized, and hovering or focusing a cycle highlights its whole column; hover a cell for the reset time and the exact recorded period. The columns stretch to the panel and scroll sideways when there are many. Charts show the latest 12 cycles, and Show all reveals up to 30. Calculation details explains the formula. The overall average divides total recorded tokens by total quota percentage points across all recorded cycles.
Recorded tokens and full-cycle estimate charts
API, Local and Device estimates stay separate. A full-cycle estimate is an extrapolation, not a fixed token allowance. Model mix, cache usage, usage on other devices and incomplete local history can change the ratio. Local recorded counts may cover only part of the account's usage.
On the Conversations tab, search by title or project and select a row for its tokens, input/cache, output, API cost, quota-window equivalents, per-model split and daily totals. Assign this conversation opens history assignment for it.
On Details, Daily breakdown by model groups usage by the selected period with a model-mix bar; expand a period to see each model. Usage by account totals Local tokens per account and shows how each row was attributed.
Refreshing reads quota and usage without sending an inference prompt. An open dashboard refreshes every minute and reuses a token-only history cache.
Select +, choose Codex or Claude Code, then choose a model and access mode. Models come from the selected installed CLI. Reasoning effort appears where supported; Codex also offers a per-project Fast mode toggle.
- Stream replies and tool activity, respond to approvals and stop an active turn from the panel.
- Paste images, drag files into the composer or use the paperclip. Preview and remove attachments before sending; attachment-only messages are supported.
- Queue follow-ups while the agent works. Edit, delete, pause or resume the queue; saved queues require Resume after reloading.
- Open History to search and resume project conversations from either provider. Edit and resend branches an earlier message while preserving the original conversation and current files.
- Restore code previews a checkpoint saved before a turn started in AgentTandem. Undo restore recovers the code from before the latest restore. External CLI messages and older history may have no checkpoint.
Project history and review shortcuts
Open Settings → Preferences to show both agents, the selected chat agent, Codex only or Claude Code only. Choose compact icons with quota, progress bars, quota rings, minimal columns (one column per window plus the lowest percentage), icons only or full text; select quota windows, reset countdowns, banked resets, account/plan details and left/right placement. Codex and Claude Code each take their own text color: keep the quota status color, pick a theme color (follows your color theme) or a custom color; low quota still shows its warning colors. With remaining quota shown, each reset countdown sits right after its window behind a clock icon (5h 75% 🕑2h14m · 7d 40% 🕑3d4h; minimal shows only the tightest window's reset); the full-text style keeps 5h reset 2h 14m. Hover retains full details in every style. ↻ 3 means three banked resets.
Status bar preferences
Open Settings → Storage to see what agent history takes up and how much cleanup has reclaimed.
The status-bar hover also shows Extension storage, its measured size and the largest data categories. Manage storage opens the extension's App data tab directly, where you can see what each kind of saved data is, who decides when it goes, and preview the exact files. AgentTandem checks its own data folder in the background every five minutes and refreshes after cleanup. Native Codex/Claude transcripts are managed separately in Conversations.
At 1 GB by default, the hover and Storage page show a warning and a notification offers Manage storage. Notifications appear at most once per day. Change AgentTandem: Storage Warning MB (agentTandem.storageWarningMB) in VS Code Settings; set it to 0 to disable size warnings. Monitoring does not delete anything.
Two headline cards stay visible on every Storage tab:
- In use now totals Codex chats, Claude Code chats and AgentTandem data, with a bar showing the share of each.
- Freed so far keeps a running total of space reclaimed by deleting conversations, restore history and unused files. It also counts the conversations, files and cleanup runs, shows the size and time of the last cleanup, and survives VS Code restarts. Only data a run actually removed is counted.
The App data tab lists each kind of data AgentTandem keeps with its size, a policy label and its main action:
- Review snapshots are cleaned automatically. Snapshots and blobs that no tracking baseline, pending change, Undo, active capture or restore point refers to, for example after you accept changes, are removed once they are at least an hour old. A pass runs shortly after VS Code starts, every few hours and after you accept or revert, never during an agent turn or other storage work, and holds the storage lock only briefly. The first run can free a lot of space, so it shows one notification with the amount and a Turn off auto-clean button. Switch it off with Clean automatically here or
agentTandem.autoCleanReviewData; Clean now runs it once, and Preview unused files… lists them first (the preview includes files from ten minutes old, the automatic pass only those over an hour). - Restore points follow a rolling Size limit for AgentTandem's data folder, 5 GB by default (
agentTandem.storageLimitGB, 0 turns it off). Above it, unused review snapshots and then the oldest restore points are removed, oldest first, until the data is back to about 90% of the limit. Pending changes, tracking baselines, the last restore backup, token usage history, accounts and attachments are never removed to meet it. Codex and Claude Code chats are not counted. - Token usage history is never deleted automatically. Compress shrinks the saved token counts without changing them (each compressed file is read back and compared before the original is removed), and Delete kept history… removes the counts of conversations whose transcripts are gone, after asking. Counts kept only because their chat was deleted are the sole copy.
- Chat attachments and Accounts and other data are always kept.
The Restore history tab filters saved restore history by project/type; select entries and Review deletion before removing them.
Review snapshots & tracking has two shortcuts, also available in the status-bar hover: Clean accepted previews unreferenced review snapshots left after Accept or other review actions, while keeping tracking baselines, checkpoints and Undo. Clean all also removes every message checkpoint and restore backup and stops change tracking across projects. Pending review and its revert Undo remain protected, so some space can remain. Both actions show the exact files and bytes before the final cleanup button. Captures less than 10 minutes old stay protected unless their saved history is explicitly removed. Workspace files, chats and recorded demos are kept.
Preview cleanup lists unused snapshots and blobs older than 10 minutes. Clean up… lets you choose what to clean — message checkpoints, restore backups and/or unused snapshots — previews exactly those, then deletes them from the same preview. Cleanup rechecks references before deleting; tracking baselines, pending review, active captures, retained checkpoints and Undo remain protected. Workspace files and native Codex/Claude conversation history stay intact. Deleting selected restore history removes the ability to restore those checkpoints.
Storage lists the projects whose saved Review comparisons and revert Undo still hold snapshot data, even when unused snapshots show 0 B. Clear saved review data… previews clearing those comparisons, Undo, checkpoints and tracking, including the affected projects and exact snapshot files. Confirming permanently removes those restore abilities while keeping workspace files and chat messages. Active captures block clearing; changed metadata requires a new preview, and newly referenced files stay protected.
Storage → Conversations scans native chat transcripts across all projects in the displayed history folders (CODEX_HOME / CLAUDE_CONFIG_DIR when configured). Last message shows the latest user/assistant message time, including subagent files for Claude; click its heading to toggle oldest/newest first. Oldest first is the default. A ~ marks file update time used when a message timestamp is unavailable. Choose 7/30/90+ days or Before date (exclusive, at local midnight), and/or a Year (local calendar year of the last message, listed with counts), then optionally filter by provider, archive status or search. Delete all matching (N) applies the filters on the host and permanently deletes every match in one run, including rows beyond Show more; it first reads any message dates still loading (N+ while they load). Delete selected immediately deletes only your selection. There is no extra confirmation step; a dialog follows the run and lists what was deleted and what was kept, with reasons. Deletion rechecks files and includes subagent payloads; a conversation with protected or newer-than-cutoff subagent history is kept instead of failing the whole run. Open conversations, queued messages and sessions updated within 10 minutes stay protected. Close other apps/terminals using these histories first. Codex archives are included; Claude Code has no native archive flag.
Native deletion is permanent and prevents resuming those conversations. Their token counts stay in the usage dashboard, as do those of transcripts a CLI removed through its own retention; Delete kept token history… in the dashboard's data coverage notes removes them after a confirmation. Restore checkpoints are managed separately. Codex uses its native thread/delete API and requires a CLI that supports it. Early Codex rollouts (before session_meta, around September 2025) cannot be read by current Codex CLIs and are not in its thread index, so thread/delete rejects them as empty; AgentTandem deletes those files directly, only while Codex has no index entry for them, after the same file and protection checks; Claude uses the bundled SDK's deleteSession. Codex .jsonl and Zstandard .jsonl.zst transcripts are supported; compressed history is streamed in memory without unpacking it to disk. Reading Zstandard requires a VS Code extension host with Node 22.15 or newer. Errors and skipped entries remain visible. Linked, corrupt or unsupported transcript stores block deletion. No cleanup runs automatically. Developer validation: npm run test:conversations:cleanup:live uses an isolated native home without copying real history or credentials; append -- --compressed to test native deletion of compressed history.
Cleanup result
Install AgentTandem in the WSL/SSH workspace so it can access that environment's files and signed-in CLIs. After switching from a local development VSIX to the Marketplace extension, uninstall agent-tandem-local.agent-tandem in that environment and run Developer: Reload Window. Keep one AgentTandem extension enabled: both builds share commands and views. Uninstalling the extension keeps its saved data.
| Shortcut | Action |
|---|---|
Enter / Shift+Enter |
Send a chat message or queue a follow-up / insert a newline |
Alt+Y / Alt+N |
Accept / reject the current change in an active review editor |
Alt+Up / Alt+Down |
Previous / next change |
Alt+PageUp / Alt+PageDown |
Previous / next changed file |
Screenshots capture the current interface with demonstration data. Account names, emails, token counts, quota values and switch logs are fictional.
See Contributing, Development and Release notes.
AgentTandem is an independent companion, not affiliated with OpenAI or Anthropic.