What Agent Park's plugin reads, keeps and sends
Agent Park's plugin gives your coding tools (Claude Code, Codex, Cursor and OpenCode, the editors and tools that take its memory tools, and any agent run through its adapter) your Agent Park workspace's memory. This page lists everything it reads on your computer, everything it keeps there, and everything it sends, from the list the plugin's own code is checked against.
1. What it never does
- It never sends your conversations, your files or your code anywhere. It sends words from your prompt to search your own memory, and counts.
- Nothing inside a project can change where the plugin sends anything, which sign-in it uses, or the text it gives your tool: those come only from Agent Park and from your own settings, never from a file in the folder you work in.
- There is no sign-in callback on your computer: the plugin never opens a port. Agent Park hands a sign-in only to the computer that holds the secret code it issued for that sign-in, after you approve it in your browser.
- It never answers a permission question for you, never blocks a prompt, and never changes what your tool runs.
- Each computer and each tool is its own sign-in, listed on the Plugins page in Agent Park, where you can remove it at any time.
2. Where it sends anything
Everything the plugin sends goes to Agent Park (app.agentpark.si), over HTTPS, as you: it signs in with your Agent Park account and never holds a key to anything else. When your computer reaches the internet through a company proxy, it goes through that proxy (HTTPS_PROXY, HTTP_PROXY and NO_PROXY are honoured), and your company's own certificates are trusted when NODE_EXTRA_CA_CERTS names them.
3. What it reads on your computer
- The prompt you type, to look up what your workspace's memory holds about it. The words it searches for are taken from your prompt; the prompt itself is not kept. Through Agent Park's adapter, the same for the text of each message you send your agent, and nothing else of the conversation.
- Your tool's own record of each request it made to its model, only to count tokens: the counts are sent (see the usage report below), the words of the conversation are not.
- The folder you work in, only to choose which of your connected workspaces this folder uses.
- Its own files under its own folder (below).
4. Every request it makes
- /oauth/device: Starts signing this computer in: the tool's name, the computer's name and its system, and the connect link or your organisation's token when there is one. (Once, when you connect.)
- /oauth/token: Collects the sign-in once you approve it in the browser, and renews it every hour. (When you connect, then hourly while you work.)
- /oauth/sibling: Signs a second tool on this computer in with the first tool's sign-in, so you approve once per computer. (When you connect a second tool.)
- /oauth/revoke: Ends this computer's sign-in. (When you uninstall or reconnect.)
- /plugin/v1/me: Checks that you are still a member of the workspace and that its memory is ready. (At the start of each session.)
- /plugin/v1/search: Searches your workspace's memory with words from your prompt, as you: you get only what you may read. (When you send a prompt, and when your tool calls the memory search.)
- /plugin/v1/artifacts: Reads one item of your memory that a search found. (When your tool asks to read one.)
- /plugin/v1/savings: Reads the figures of what memory saved, for your tool to tell you. (When your tool asks.)
- /plugin/v1/remember: Saves a note you or your tool chose to keep, and, only if you turned session notes on, a short note of what the session worked on. (When asked to save, and at the end of a turn with session notes on.)
- /plugin/v1/usage/observations: The usage report: your tool's own token counts for each request it made (fields below). Never a word of the conversation. (At the end of a turn, and at the next session start for anything not yet sent.)
- /plugin/v1/usage: Reads your own figures back, for /agentpark:usage and the status line. (When you ask, and once a day at a session's start.)
- /plugin/v1/plan: The plan report: how much of your plan's 5-hour and weekly use your tool says is used (fields below). (At the end of a turn.)
- /plugin/bundle/latest: Asks which version of the plugin is current. (When you connect or update.)
- /plugin/bundle/ticket: Asks for a one-time download ticket. (When you connect or update.)
- /plugin/bundle: Downloads the plugin, checked against the SHA-256 Agent Park names before it is used. (When you connect or update.)
Its code also holds 7 routes that a tool connected through Agent Park never calls: the tools behind them are not offered there.
5. What it keeps on your computer
- ~/.agentpark/credentials.json: Claude Code's sign-in: an access token that lasts an hour and the token that renews it, for each workspace you connected. Readable only by you (mode 0600).
- ~/.agentpark/<tool>-credentials.json: The same for Codex, Cursor and OpenCode, each editor you connect and Agent Park's adapter, each its own sign-in.
- ~/.agentpark/plugin-data/: Claude Code's records, per session: what was recalled into the conversation (so nothing is given twice), the token counts waiting to be sent, whether this session's check passed, and any report that could not be sent yet. No conversation text.
- ~/.config/agentpark/plugin-data/: The same records for Codex, Cursor, OpenCode, the editors and the adapter (in the folder XDG_CONFIG_HOME names, when it is set), and, for an editor's memory tools, when they last opened the page to approve this computer (at most once each ten minutes).
- ~/.agentpark/marketplace/: The plugin itself, as downloaded from Agent Park.
- ~/.agentpark/installed.json: Every file the plugin put in a tool's folder and every setting it added, each with its SHA-256, so uninstalling removes exactly those and nothing of yours.
- ~/.agentpark/connect/connect.mjs: The connector, checked against the SHA-256 the connect line pins.
- ~/.cursor/plugins/local/agentpark/: Cursor's copy of the plugin (Cursor only).
- ~/.config/opencode/plugins/agentpark.js: OpenCode's plugin file, and an "agentpark" entry under "mcp" in OpenCode's settings file for the memory tools (OpenCode only; a settings file that does not parse is never rewritten).
- Claude Code's and Codex's own plugin folders: The plugin as each tool installs it, through the tool's own commands.
- An editor's own settings file: Only when you run its connect line with --write: an "agentpark" entry for the memory tools (and, for Zed and JetBrains, two agents run through Agent Park's adapter), between marker comments where the file takes comments, and nothing else of the file changed. A file that does not parse is never rewritten; uninstalling removes exactly what was added.
6. The usage report, field by field
- source: Which tool's own counts these are.
- provider: Which kind of service your tool sent the request to, for pricing the counts.
- model: Which model your tool used, as your tool names it, for pricing the counts.
- conversationKey: The tool and its own session id, and the folder's name with a short hash, so a figure is put on its session.
- requestId: Your tool's id for the request, so a report sent twice is counted once.
- providerRequestId: Always empty from the plugin.
- occurredAt: When the request was made.
- arm: Empty, except during a measurement you agreed to.
- usage: The token counts: input, read from cache, written to cache, output and reasoning, as your tool counted them. A count your tool did not give is sent as unknown, never as zero.
- status: Always empty from the plugin.
- outcome: That the request was answered.
- latencyMs: Always empty from the plugin.
- gatewayMs: Always empty from the plugin.
7. The plan report, field by field
- windows: For your plan's 5-hour and weekly windows: which window, the share used, and when it resets, as Claude Code reports them.
- at: When Claude Code reported them.
8. The records it keeps
- recall: Each recall: when, and whether it found something, nothing, or failed. Not what it found.
- usage: This computer's copy of the counts it reported.
- codex-usage: Codex's counts per request, until they are sent.
- acp-usage: Through Agent Park's adapter, an agent's own counts for each turn (and the cost it reported), for an agent none of the plugin's hooks run in, until they are sent.
- ruleset: Each time the working conventions were handed to your tool, with the text's version.
- folds: Only when you turn folding on: each long command output it folded and how much smaller it got.
- gates: Only when the evidence check is on: when it asked your tool to open the lines it relied on.
- dedup: Only when you turn on storing repeated output once: which command outputs it shortened, and by how much.
- repo-map: Only when you turn on the repository map: that a map of the project's file names was given, and its size. The map itself is made on your computer from your project and is not sent to Agent Park.