Getting started
Read-only context
Read sources without giving the agent write authority.
Source articles can contain malicious instructions alongside useful facts. For research-only runs, a trusted client can restrict the agent's MCP connection to reading core context. The server blocks changes, sharing and credential issuance.
Available context
Read and search skills, context objects, wiki articles and memory under current access rights. New articles in a shared wiki remain available. Revoked sharing is checked on subsequent requests.
Writing, deleting, sharing, account administration and domain plugins are unavailable in this mode. Workspaces select content; they are not security boundaries. Org API keys are not guaranteed read-only credentials.
Client integration
- The trusted controller authenticates with personal OAuth.
- Outside the model's tools, call
POST /mcp/context-read-grantswith the OAuth bearer and a JSON body such as{"expiresIn":900}. - The response returns
credentialandidonce. Give only this restricted bearer to the agent's MCP transport athttps://meinkontext.de/mcp. Never print credentials in prompts, tool results or logs. - Using the original OAuth bearer, revoke through
DELETE /mcp/execution-grants/{id}after the run, including failures.
Lifetime is 60–3,600 seconds, default 900. The credential cannot renew or widen itself. Expiry is an access failure; never silently fall back to full authority.
Client support and limits
Ordinary OAuth connections are not automatically restricted. Clients must implement this separate authority flow; there is no URL switch or universal Codex, Claude or Hermes setting. No local companion service is required; all context delivery remains MCP.
Keep full OAuth inaccessible to the model, including other connections and shell tools. Read-only authority prevents changes through this connection, not incorrect answers or disclosure of readable content through other client tools.