meinKontextDocs
Navigation & search

Getting started

Accounts, apps and agents

Agents under your account with selected context, their own tasks and controlled integrations.

Your knowledge should not need a new home whenever another AI joins in. What matters is who owns the context and which permissions a connection uses.

What works today

Your goalCurrent approach
More apps should use selected context.Choose an agent profile or explicitly your personal account during OAuth authorization.
An agent should independently own its context.Use a separate account and share selected content through buckets.
A supported client should only read core context.A trusted controller can issue a short-lived read-only credential. This is not an independent agent identity or a way to select individual private facts.

You do not need a new account for every app. Personal OAuth and managed agents are different access types: a managed agent acts under its own identity.

Connect an app · Headless sign-in · Read-only context

Separate agent access

Manage named agents under your existing account in Agents. Context stays with its owner; each agent gets a distinct identity and explicit permissions.

You manageBenefit
Name and identityRecognize which agent performed a change.
Permitted areas and actionsEach agent works only within its granted scope.
Lifetime and individual credentialsRotate or revoke one credential without setting up every agent again.
Human approvalsAn agent's work is not automatically presented as your personal decision.

For example: a research agent reads a shared wiki while a work agent edits tasks. Both work under your account with different rights. No copied knowledge or additional owner account is needed. Generative AI and agent execution stay outside meinKontext; only search currently uses a small local model inside meinKontext, as a temporary test.

Set up and share

  1. Open Agents under Context. The page lists all your agents with their state (Connected, Not connected, Revoked). Use New agent to create one, for example “Moss”; the new agent opens right away.
  2. Click Add access. Find a wiki, bucket, object, optional workspace or plugin content such as tasks. You can select several content entries for read access. For plugin content, choose its access level; the row names the plugin it belongs to.
  3. Use Connect the first time. Save the once-only credential in your MCP client's secret store. Never put it in chat or context documents.

The agent's access now appears under Access. To open another agent, choose its row in the list; its details appear beside it, on a phone in place of the list, and Back to list returns. Each row shows its current access on the right. For core content choose Read, Read · Writing needs approval, or Read and write. Roles on plugin content can be changed through the permission menu. The separate Remove action revokes that grant. The list has no expandable levels. Sharing a wiki includes its future articles and original files. Every new agent receives its own owner-held home workspace, alongside any additional shares.

For separate tasks, share the tasks of a workspace such as “Moss” under your account with write access. For external actions, select Integrations, then choose Edit on the connection to set each action: Not allowed, After my approval or Allow directly. This requires Integrations version 0.0.2. Without an action grant, the agent cannot use that connection.

Connect your agent

  1. Open your agent under Agents and select Connect in its Connection section (later Manage connection, also in the agent's “⋮” menu).
  2. Choose a lifetime and click Issue new credential. Save the once-only MCP bearer in your client's secret store.
  3. Add an MCP connection in your client with these settings:
SettingValue
Server URLhttps://meinkontext.de/mcp
TransportStreamable HTTP
AuthenticationBearer token: the agent credential you just issued

If the client expects custom HTTP headers instead of a token field, use Authorization: Bearer <agent-credential>. Replace the placeholder with the credential; a dedicated bearer-token field takes only the token. Never store it in chat or a repository. The server URL above belongs to this instance: preview credentials work only in the preview.

  1. Enable the connection or reconnect the client. The credential automatically identifies the agent and its grants; no additional agent or bucket ID is needed in the URL.
  2. Use Check access in the agent's “⋮” menu for that credential, then test that the selected content is reachable in the client.

Connect with OAuth, for example from Luma

  1. Create the agent under Agents and choose its access.
  2. Start an MCP connection to https://meinkontext.de/mcp in your app and sign in to meinKontext.
  3. Under Profile and access, choose the agent. Full profile includes its full profile; an individual listed workspace restricts this connection further.
  4. Click Allow. The app receives a separate credential for this profile, without copying a token.
  5. Check in the client that shared content is reachable and private content outside the selection is absent.

My account · full personal access connects the app as you. Choose it only when intended. Existing personal connections do not become agent connections automatically: disconnect them in the client and authorize again with the intended profile.

Token refresh preserves the selection. Additional grants require new authorization; revocation also blocks otherwise valid tokens. Revoke the relevant OAuth credential under Manage connection. Each client gets its own work session. OAuth agent connections expire after 30 days; connect again afterwards. A workspace selection never grants access beyond the profile.

See Connect your app for client setup. Luma needs to open the OAuth authorization page. Profile selection happens in meinKontext, not in an additional Luma settings field.

Unlimited issues a bearer credential without an expiry date. It remains revocable and gains no additional permissions. The default is 24 hours. For an existing agent without a home, choose Set up agent home and issue a new credential afterwards.

After expiry, issue a new credential here and replace the token in the client. Do the same when you want to include additional grants. Each concurrent worker gets a separate credential.

Day to day

  • Add access adds to the list. The search covers your whole account as you type: content and plugin content such as tasks. Selection opens in its own dialog.
  • Waiting for you appears only when approval requests are pending and links to the original request.
  • Connect appears when no valid credential includes all saved access. New rights require a new credential; existing credentials never grow automatically. The same holds after a plugin update: a credential is valid for the plugin version it was issued with, and the agent needs a new one afterwards. Removal takes effect immediately for new requests.

What waits for you also appears under Action items on the overview and in list_action_items: every change an agent proposes, and every agent whose new access takes effect only after connecting. An entry leads straight to the review dialog or to Manage connection and disappears once you have decided or connected.

The Manage connection dialog holds setup and individual credentials. The agent's “⋮” menu also offers Check access and, after a confirmation, Revoke agent. Check access shows what a selected credential can actually reach. This snapshot never runs a provider action; every actual request checks access again.

A manually issued credential lasts 24 hours, seven days, 30 days or until revoked without a time limit. OAuth agent credentials last 30 days. Each concurrent worker needs its own credential; reusing a credential continues the same work session. Already-read copies cannot be recalled. If a stopped agent holds a task lease, the owner of those tasks can use Recover this lease with a reason.

What a grant includes

Select content individually or as a whole (a whole wiki, bucket or workspace). Only explicitly shared core content appears in discovery, search and file reads. An optional workspace selection limits this agent credential to its current contents; ordinary personal workspace connections remain catalog views.

Tasks, Calendar and Integrations support service identities. Messages and plugins made in the plugin builder remain unavailable until their identity contracts support them. Installation or administration does not grant content rights. Your personal memory stays private until you explicitly share individual facts. You must own a selected bucket; received buckets cannot be delegated onward here.

Give an agent access to integrations

Choose a rule for each operation on each selected connection:

  • Not allowed: The agent cannot invoke the operation.
  • After my approval: The request waits for your browser review of its stored content. This can also be required for reads.
  • Allow directly: The agent may invoke the operation without another confirmation while current rights and its original credential ceiling remain valid. This can explicitly include writes.

Review shows the connection, operation, complete arguments, requesting agent and work session. You approve the original request or reject it. A rejected request remains blocked on replay; the agent cannot approve itself. Review expires after fifteen minutes, or earlier when the credential or authority expires. Changes to the connection, tool definition or permissions can require a new request. Repeated approval of the same request does not dispatch it again. Check uncertain outcomes at the provider before another attempt.

Provider credentials remain with the protected executor. The agent does not need access to your Secrets to use a connection you configured.

Your agent’s memory

Each agent stores its own thoughts and facts in its own workspace under your account. Facts with the same name in your personal memory or another agent’s workspace remain separate.

In Memory, open the workspace picker beside the page title. Search for the agent by name and select its area. All workspaces shows the facts of every workspace again. You can search its facts, inspect the fact map and history, forget facts or permanently delete them. Your account export keeps workspace memories separate.

Use Agents → Add access to share individual facts or an entire additional workspace. Workspace access includes facts saved there later. An individual fact grant follows revisions; forgetting and recreating it creates a new fact without the previous grant. Hidden origin nodes and links stay invisible to the agent.

Facts also support Read · Writing needs approval. Review the previous content and proposal, then allow this change, further changes to that exact fact for this connection, or until revoked. Approving one fact does not authorize changes to other facts in the workspace.

Reconnect after adding access. Older credentials do not automatically gain memory rights. In the connected app, the agent can read through recall and request changes through remember or forget. New personal agent facts default to its home workspace. After you approve a pending request, the agent must retry the same operation.

Moss under your account

Moss receives a home workspace under your account. Its own files plus additional shares form its profile; you can also grant dedicated tasks. Issuing a credential does not move data from Moss' existing account. Tasks, historical names and existing clients need separate mapping and verification before cutover. Retire the old account only after verified adoption.

Does an independent agent account still make sense?

Yes, when the agent should own its context independently. It can use shared skills, whole wikis and plugin content such as tasks while maintaining private content. Revoking a share preserves its own content. Existing agent accounts will not be merged automatically.

Can an agent decide on my behalf?

A managed credential identifies the agent itself. Write permission does not prove that you personally approved something. Narrow, verified delegation would be a separate later step. That requires an explicitly human-operated approval path. For existing connections, an AI using your personal OAuth acts as you; that login alone does not prove human review.

How will this be verified?

The acceptance checks real MCP connections for separate rights, revocation affecting new operations, accurate attribution, resistance to forged human approvals and consistent enforcement on search, export and retries. Copies already read cannot be recalled.

The separate “Agenten-Zugänge und Urheberschaft” suite tests attributed findings, human decisions, concurrent sessions and withdrawn write access. Langfuse records model runs and individual checks. Technical authorization rules are tested independently; absent or incomplete model runs never count as passing.

SDK contract and acceptance · Existing test methods

Own workspace and additional access

Find agents under Context → Agents. Create profiles and manage their access here. Apps & API keys, in the account menu at the foot of the navigation, lists connected clients and their credentials.

The home workspace uses the agent name in kebab case, such as agent-moss or agent-research-bot. If the name is taken, it receives -2, -3, and so on. Existing automatically generated UUID names also receive readable names; their previous MCP addresses remain valid aliases. Contents and permissions stay with the same workspace.

Every newly created agent receives a home workspace owned by your account. It can hold instructions, skills and notes, and it covers the plugin content lying in it: tasks or a wiki in the agent's workspace needs no separate share, with write access under Read and write and read access otherwise. The Agent home row names the plugins it reaches this way. Revoking the agent keeps these owner-held contents. An active agent's home cannot be deleted; once you have revoked the agent, deleting its workspace removes the revoked agent with it. Existing credentials never gain access automatically; the plugin content of the home needs a credential issued after this change.

The agent gets its definition when it connects. The instructions it receives name the instructions, skills and other objects you shared with it, and its MCP client lists the shared skills through the standard skills surface (skills/list, skills/get, skill://), so it needs no meinKontext-specific call to load them. Nothing outside its access appears there.

With Read and write on its home, an agent can also create buckets of its own there (create_agent_bucket), link objects of its home into them (add_to_agent_bucket) and list them (list_agent_buckets). Such a bucket belongs to you and lives in the agent's home workspace. Only you add members; the agent cannot invite anyone or reach a bucket outside its home.

A profile identifies the acting owner or agent and its complete permissions. A workspace is an optional narrower view. An agent normally sees its home and additional shares together. ?workspace=<name> restricts a connection to a whole workspace already included in its credential. An unavailable workspace is rejected.

?profile=service:<agent-id> can assert the identity bound to a credential. It cannot switch identities or grant access. It does not replace profile selection in a client-initiated OAuth authorization flow.

Review file writes

Choose Read · Writing needs approval to allow reading while requiring approval for file changes. Open Review change on the agent. Compare the previous content with the proposed content, select files, then choose:

  • Only this change: the exact stored proposal; a different proposal needs fresh approval.
  • For this connection: later changes to these exact files using the same agent credential. Other connections need their own approval.
  • Until revoked: later changes to these exact files across this agent's credentials, while the parent access remains valid.

Use notes/*.md to select Markdown files directly in that folder, or notes/** to include its subfolders. The dialog lists the actual selected files. Future files are never included automatically.

After approval the agent retries the original request. Identical retries do not write twice. If the published revision changed in the meantime, the agent must read it again and prepare a fresh proposal. Rejecting prevents that proposal from running.

Write permission cannot issue credentials or approve the agent's own requests. Producer-managed wiki articles continue to use their producer's write path. Plugin and external-provider actions retain their independent authorization and approval rules.

Protect definitions in the existing home

An existing object grant set to Read · Writing needs approval also restricts direct write access through the home or a bucket. Read on the object grant denies writes to that object. Other working files in the same home retain their own policy. A restriction captured in an issued credential remains its ceiling; later grants do not widen it.

The agent calls write_agent_files with the current revision. For a pending result, review the existing file proposal. After approval, the agent repeats the identical call with the same operationId; that completion publishes the change. Access revoked in the meantime is checked again. Approval alone does not publish.

Poll version changes

list_workspace_objects returns an agent's authorized stable versions in a fully granted owner workspace. workspace accepts its ID or name; kind filters the object kind. The response contains kind, slug, seq, version_id, publisher, and an etag, without file bodies.

Pass the previous tag as if_none_match to receive unchanged: true when the listing is identical. Live access is checked even then. Plugin records and versions pinned away from stable are excluded. publisher identifies the evidenced publishing actor, not a human approval. If the publication event is absent from retained audit history, its value is null.

Manage agents over MCP

Your personal MCP connection reaches the same steps as the Agents page through these tools:

On the pageMCP tool
Agent list with state and credentialsagent_access_list lists your agents, non-secret credential references and the plugin content with the roles you can grant.
New agentagent_access_create with name creates the agent and its home workspace, which the agent may read and write. It shares none of your other content and issues no credential.
Access level on plugin content, Removeagent_access_grant with agent_id, collection_id and a role_id from agent_access_list; role_id: null removes the grant.
Check accessagent_access_list with both agent_id and credential_id returns the live preview of that credential. Both parameters are required together; the call changes nothing.
Preview before a changeaccess_preview dry-runs a grant change for one of your agents and lists what it would gain or lose.
Revoke agent, revoke one credentialagent_access_revoke with exactly one of agent_id or credential_id.

Two steps stay outside MCP. Connect issues a credential: use the button on the page, or let a trusted controller post to /mcp/agent-credentials (SDK contract); no tool result ever contains a credential. Wikis, buckets, objects, workspaces and facts are shared under Add access on the page. MCP never widens what an existing credential reaches: access that agent_access_grant adds beyond a credential's original ceiling takes effect only with a credential issued afterwards outside MCP. Removing or lowering a role applies to existing credentials at once.

These tools answer only on your personal connection. Agent, API key and execution connections neither list nor call them.