Appearance
MCP
geslar mcp serve is a stdio MCP server over an agent profile. Any MCP client that can start a local stdio server can use it; geslar mcp init configures Claude Code, Claude Desktop and Cursor for you, and prints the entry for any other client (Other MCP clients):
sh
geslar mcp init ci-bot # Claude Code (default; --scope user or local)
geslar mcp init ci-bot --client cursor # claude-code, claude-desktop, cursor, or json (print only)- The entry is only
{ command, args, env: { GESLAR_AGENT: "ci-bot" } }, never the token. It startsgeslar mcp serveifgeslaris on your PATH, otherwisenpx -y @geslar/cli@<exact version> mcp serve. - A configuration file that cannot be parsed (for example one with comments), or is a symbolic link, is not touched: the entry is printed for you to paste. Otherwise a backup is written first, other servers are kept, and running it again changes nothing.
- For Claude Code,
mcp initrunsclaude mcp add-jsonwhere it can and otherwise prints the exact command. Claude Desktop on Windows is not yet verified on a real client: the entry for it usescmd /c npx …whengeslaris not on the PATH. Check that the server shows as running and report problems. mcp serveneeds the OS keychain (no key file) and refuses to start whileGESLAR_SERVICE_TOKENis set.
No hosted version
There is no hosted version of the Geslar MCP server, by design: your secrets are decrypted only on your own computer, so the server has to run there. Chat apps that run only in the browser or in the cloud cannot start it; use a client that runs on your computer.
The five tools
| Tool | What it does |
|---|---|
geslar_list_vaults | Lists the Vaults this agent profile can access (id and name). |
geslar_list_items | Lists the names and types of the items in one Vault, with the standard field names each type has. |
geslar_whoami | Shows this agent profile's name, expiry, the Vaults it can access and its active approvals. |
geslar_provision_env | Adds KEY=geslar://… reference lines to an env file (.env, .env.*, *.env) under the server's root, never values. An existing key is a conflict. |
geslar_run | Starts a program (command and arguments, no shell) with secrets in its environment. A person must have allowed the command and approved access to the secrets. |
No tool returns a secret value by design. The one way a program's output can reach the model is --return-stdout-tail, below.
geslar_run answers only with the exit code, how the program ended, the duration and the number of output lines. The model gets no output unless you start the server with --return-stdout-tail (a filtered tail, best effort, not a security boundary; values shorter than 6 characters are never masked in it). References are accepted in env and env_file only, never in the command, its arguments or the directory.
mcp serve options: --agent <name> (or GESLAR_AGENT), --root <dir> (root directory for geslar_provision_env; default: the working directory), --pass-env <names> (names of variables of this environment the child may inherit, comma-separated, for example HTTPS_PROXY,SSL_CERT_FILE) and --return-stdout-tail.
Which commands the agent may start
You decide which commands the agent may start, in a terminal, with a typed confirmation:
sh
geslar mcp allow-cmd "npm test" # lasts 1 hour by default (--ttl up to 8h; --max-uses n)
geslar mcp grants # active patterns
geslar mcp revoke --all # or: geslar mcp revoke <id or pattern>A pattern lasts one hour unless you set --ttl (at most 8 hours), and --max-uses limits how many times it can be used. Keep patterns narrow (npm test, not npm *).
A pattern is not a security boundary. A command that can run arbitrary code (a shell, node, python, npx and the like) can read the secrets in its own environment and send them anywhere. What protects your Vault is the Vault you gave the agent, its time and read limits, and the approval by a person with 2FA, not the pattern.
Other MCP clients
Any MCP client that can start a local stdio server should work: geslar mcp serve speaks the MCP protocol over standard input and output only (no network port) and relies on nothing that is specific to one client. The configuration of a client names a command, its arguments and an environment; for Geslar they are:
- command
geslar, argumentsmcp serve, environmentGESLAR_AGENT=<profile name>(instead of the environment variable you can add--agent <profile name>to the arguments); - when
geslaris not on the PATH: commandnpx, arguments-y @geslar/cli@<exact version> mcp serve, the same environment.
The profile must exist (geslar agent add <profile name>) and the OS keychain must be usable: the server never asks for a password. mcp init can print the entry in the shape that many clients read:
sh
geslar mcp init ci-bot --client jsonjson
{
"mcpServers": {
"geslar": {
"command": "geslar",
"args": [
"mcp",
"serve"
],
"env": {
"GESLAR_AGENT": "ci-bot"
}
}
}
}Where that goes and what the fields are called is in the documentation of your client; mcp init does not guess it. On Windows the entry that mcp init writes for Claude Desktop starts npx through cmd (cmd /c npx -y …), which is not yet verified on a real client; a client that starts commands without a shell may need the same, and installing geslar globally avoids the question.
Tested with: the official MCP TypeScript SDK client (pinned in the CLI, run in its test suite against the real mcp serve over a stdio pipe and a fake Geslar server) and MCP Inspector 2.9.0 in CLI mode (initialize, tools/list with the five tools, and calls of geslar_whoami and geslar_list_vaults, against a fake Geslar server). Both check the protocol; neither is a client that people use. mcp init configures: Claude Code, Claude Desktop and Cursor, and prints the generic entry above for anything else. If your client works, or does not, tell us which one and which version in an issue.