Security
What the Sypnia MCP server can and cannot do, the one tool that writes and how to switch it off, and what to know before you connect an agent to production data.
Engine Cloud Read this before you connect an agent to a project with real users. It says exactly what the server can do with the token you give it.
What the server can and cannot do
| It can | It cannot |
|---|---|
| Read problems, their analysis, stacks, traces and releases | Change your code, your repository or your deploys. It has no access to them |
| Read sessions as a list of events: pages, clicks by their label, failed calls, errors | Return what a person typed. It is never collected |
| Read performance numbers, uptime status, and search | Return replay recordings or page snapshots. No tool does, with any token |
| Mark a problem as resolved (one tool, see below) | Ignore or delete a problem, delete a session or a person’s data |
| Create, read or revoke keys and tokens; change settings, alert rules, webhooks or uptime checks | |
| Open, merge or close a pull request | |
| Reach another project. A token is one project |
Nine of the ten tools only read. Reading has no side effect: asking twice changes nothing and costs nothing but a request.
The one tool that writes
resolve_problem marks a problem as resolved. It is there so the loop can close: the agent fixes the code, you merge, the agent marks the problem.
It is safe by design:
- It is reversible. Reopen the problem at any time, in the app, the Console, the CLI or the API.
- It undoes itself. If the failure happens again, Sypnia reopens the problem and creates an alert. A problem resolved too early cannot hide a failure that is still happening.
- It is recorded. The problem keeps who resolved it (
resolved_by: mcp:<token name>), when, and the note. - It is the only one. No other tool changes anything.
The tool’s own description tells the agent to call it only after the fix is merged or deployed. Most clients also ask you before a tool that writes runs.
Make the server read-only
Remote server
Add ?read_only=true to the address. The tool is then not listed at all.
claude mcp add --transport http sypnia \
"https://mcp.sypnia.com/mcp?read_only=true" \
--header "Authorization: Bearer YOUR_API_TOKEN"In a JSON configuration: "url": "https://mcp.sypnia.com/mcp?read_only=true".
Local server
Pass --read-only, or set SYPNIA_MCP_READ_ONLY=true.
claude mcp add --env SYPNIA_API_TOKEN=YOUR_API_TOKEN --transport stdio sypnia \
-- npx -y @sypnia/mcp --read-onlyWhole engine
On a self-hosted engine, set SYPNIA_MCP_READ_ONLY=true to remove the tool for every connection, whatever the client asks. SYPNIA_MCP_ENABLED=false removes the MCP server altogether. See the configuration reference.
Read-only here is about the MCP server. The same read token could still resolve a problem through the HTTP API, because resolving is one of the reversible marks that scope allows.
Text from your app reaches your agent
Tool results contain text that Sypnia did not write: error messages, the labels of things people clicked, page titles, names your app sent with identify(). A person using your app could put words there on purpose, hoping your agent reads them as instructions («ignore your instructions and…»).
What Sypnia does about it:
- The server’s instructions tell the agent that this text is data and must never be followed.
- Results are structured: such text sits in named fields (
message,label,text), never mixed into the server’s own wording. - Strings are cut to their limits (an error message to 500 characters, a label to 80), so there is little room.
- Nothing a person typed into a form is ever in a result.
What you should do: keep your client’s confirmation for commands and file changes switched on when the agent is working from production data, and review what it changes. This is true of any tool that brings outside text to an agent, not only of Sypnia.
Where the data goes
The server gives results to your coding agent. Your agent sends them to its model provider, under your agreement with that provider. Sypnia does not choose that model and does not see that conversation.
So what a tool returns is what leaves: code locations and source lines of your app, error messages, click labels, the ids of people, and their names and e-mails only if your app sends them with identify(). If that must not reach your agent’s provider, do not send names and e-mails to Sypnia, or use a token for a project that has none.
The token
- Use an API token with only
read. See Authentication. - One token per agent and per person, named after them.
- Never an
admintoken, never the root token. - Read it from the environment; do not commit it.
- Revoke it when a laptop is lost or a person leaves: it stops working within a minute.
Remote or local
| Remote server | Local server (@sypnia/mcp) | |
|---|---|---|
| Who runs it | Sypnia (Cloud) or your engine | Your machine |
| What it holds | Nothing. Each call goes to the engine with your token | Nothing. It forwards each call to the API |
| What you install | Nothing | A package from npm, MIT licence, open source |
| The token is in | Your client’s configuration, sent over HTTPS | Your environment |
The local server makes no other network call than to the API address you set.
What is logged
For each tool call the engine logs: the time, the token’s id, the project, the tool’s name, whether it succeeded and how long it took. It does not log the arguments or the result. On Sypnia Cloud these logs are kept for 30 days.
Limits
- 600 requests a minute per token, shared with any other use of that token. One tool call is one to three requests.
- A result is at most 60,000 characters.
- A tool call that takes longer than 30 seconds is stopped and answers an error.
Report a problem
If you find a way to make the server do something this page says it cannot, write to the address on Security practices.