Context across tools

AgentKeyRoute Context

One library of snippets, saved AI replies and projects. Syncs across the browser extension, Claude Desktop, mobile Claude, Cursor and anything else that speaks MCP.

What you get

AgentKeyRoute Context is a personal library of context items you can call into any AI chat. Items can be short snippets ("my stack", "writing voice"), saved replies you want to reuse, files from an AgentKeyRoute workspace, or whole projects that group them. The same library is available from three places:

  • The AgentKeyRoute Context browser extension — side panel, right-click, keyboard shortcuts on claude.ai, chatgpt.com, gemini.google.com and ~10 other chat sites.
  • Any MCP-aware client — Claude Desktop, mobile Claude, Cursor, Zed, VS Code — by adding one MCP server URL and a device token.
  • Directly, through the REST API, if you're building your own tool.

The extension and the MCP server both call the same endpoints, so a snippet saved in one place shows up everywhere else within a sync cycle.

Download the browser extension

AgentKeyRoute Carry v0.1.9 — load unpacked while we're working through Chrome Web Store review.

⬇ agentkeyroute-carry-0.1.9.zip — SHA-256

Unzip, then in Chrome open chrome://extensions, turn on Developer mode, and click Load unpacked on the unzipped folder.

Mint a device token

Every non-browser client (Claude Desktop, mobile, Cursor…) authenticates with a bearer token called a context token. You mint one per device, name it, and paste it into the client once. Tokens are user-scoped: they can read and write your library, and nothing else, regardless of what else your AgentKeyRoute account can do.

Create a token (POST /api/context/tokens)
curl -sS https://app.agentkeyroute.com/api/context/tokens \
  -H "Content-Type: application/json" \
  -H "X-Requested-With: covenra" \
  --cookie "cov_session=$COV_SESSION" \
  -d '{"label":"Claude Desktop on MacBook","client_kind":"mcp"}'

The response includes a one-time token field of the form cvctx_…. Copy it immediately — only its hash is stored, so if you lose it you have to mint another.

Connect Claude Desktop

Add a new MCP server to Claude Desktop's config file:

# ~/Library/Application Support/Claude/claude_desktop_config.json
{
  "mcpServers": {
    "covenra-context": {
      "url": "https://app.agentkeyroute.com/api/mcp/context",
      "headers": {
        "Authorization": "Bearer cvctx_YOUR_TOKEN_HERE"
      }
    }
  }
}

Restart Claude. You'll see covenra-context in the tool list; the model can now call list_items, search_items, get_item, add_item, update_item and delete_item. Try asking "Save this reply to my AgentKeyRoute Context as 'answer about X'".

Connect mobile Claude

Mobile Claude supports remote MCP servers through the same URL. In the app, open Settings → Connectors → Add custom, and paste:

FieldValue
URLhttps://app.agentkeyroute.com/api/mcp/context
AuthBearer
Tokencvctx_… (your device token)

Your library is now available from any mobile conversation, same tools as on desktop.

Connect Cursor, Zed, VS Code, anything else

Any MCP-compatible host works the same way. Point it at the URL above with a Bearer header. The server advertises itself at GET /api/mcp/context so a client can verify it's reachable before sending JSON-RPC.

What the tools do

ToolPurposeArguments
list_itemsList items (title + preview).kind?, limit?
search_itemsSubstring search across title and body.query, limit?
get_itemFetch one item's full body.id
add_itemCreate a new item.body, title?, kind?, tags?
update_itemPatch an item (only supplied fields).id, and any field
delete_itemSoft-delete an item.id

Items also appear as MCP resources at covenra-context://item/<id>, so a client that prefers resources over tools can list and read them that way.

REST API

If you're building your own tool rather than wiring up an MCP host, the same data is available at these endpoints. Authentication is either a session cookie (web, the extension) or a bearer cvctx_… token in the Authorization header.

EndpointDoes
GET /api/context/itemsList items. Add ?since=<RFC3339> to pull changes (including soft deletes) since a timestamp.
POST /api/context/itemsCreate an item.
PATCH /api/context/items/{id}Update an item.
DELETE /api/context/items/{id}Soft-delete an item.
GET /api/context/usageCurrent count and byte total, plus your plan limits.
GET /api/context/tokensList device tokens.
POST /api/context/tokensMint a token (returned only once).
POST /api/context/tokens/{id}/revokeRevoke a token immediately.
POST /api/mcp/contextMCP JSON-RPC endpoint — described above.

Sync loop

Clients pull with ?since=<ts>, then save the response's server_time for the next call. A deletion comes back as an item with its deleted_at set, so local caches know to drop it.

Plan limits

PlanItemsStoragePriority sync
Personal (free)500100 MBNo
Personal Pro ($4/mo)Unlimited2 GBYes
Team and aboveUnlimited5 GB per userYes

The browser extension shows your current usage in its side panel; the REST endpoint GET /api/context/usage returns the same numbers for other clients.

Troubleshooting

SymptomFix
401 Unauthorized from MCPThe token was mistyped or revoked. Mint a fresh one and paste it again. Tokens starting with anything other than cvctx_ aren't ours.
Tools don't appear in Claude DesktopConfirm the JSON parses (a trailing comma breaks it silently) and restart Claude fully, not just the window.
Changes don't appear on another deviceBoth clients pull with ?since=. If a client hasn't synced in a while, pull without ?since= once to reset its state.
items_quota / storage_quota error on addYou've hit your plan limit. Delete old items or upgrade to Personal Pro.

Security notes

  • Context tokens grant only context.* capabilities. They can't read or modify anything else your AgentKeyRoute account is attached to — even if you're the Owner of an organisation.
  • Only the token's SHA-256 hash is stored. We can't show you a token again after it was minted.
  • Revoke a lost token at POST /api/context/tokens/{id}/revoke or in the console. Revocation takes effect on the next request.
  • Rows are row-level-isolated to your tenant in Postgres; a query missing a user_id filter would still never see another tenant's data.