Skip to main content
Manage services from meerkat-connect in the selected realm or through your application’s integrated Console. Application authors can also use the saved-request and connection-view APIs below.

​
Ordinary CLI and Console setup

The two product paths have different destinations:
  • CLI: choose the intended realm, follow the setup agent’s terminal questions and review the installation and requested access. After confirmed setup, use the integration from ordinary agents in that realm under current permissions.
  • Console: begin with the original task. The integrated setup panel presents questions, protected human actions and exact reviews. Successful setup returns the capability to that same task through its native owner.
The CLI archive includes its setup runtime. In an integrated Console, the application owns login, setup lifecycle and the original conversation. Follow Request an integration for the guided journey and Authentication for protected human actions. Keep installation, saved credentials and task grants distinct; use readiness to interpret actual outcomes. A saved definition or accepted decision alone does not establish a usable connection.

​
Advanced library and archive workflows

The current lower-level local setup tools can save exact requests, reopen them for readback, or propose their settings in a new setup. Use this section when operating the advanced library setup, whose host preparation and recovery are explicit responsibilities.

​
Save setup work

In setup history, select each relevant job with Add to recovery file, then choose Save recovery file. The result is toolkit-recovery.json. An agent can obtain the same file with setup_recovery_manifest:
Replace the placeholder with an actual selected job ID. Save the returned text as UTF-8 unchanged and retain its sha256; the tool itself writes no file. Jobs need a durably retained request. An already anchored running job can be saved, so export does not prove completion or stop its worker. Keep the original workspace and installed/acquired files at their existing paths. The file contains request references, not software, credentials or the original agent’s runtime. Review its contents before sharing it with another person or setup agent.

​
Read saved work after process loss

A generated setup command owns a single-use workspace. Restarting that command against the same directory is refused. First establish that the old workers have stopped and retain exclusive custody of their files; recovery does not find or terminate those processes. Use the library host-preparation command with a new configuration output and future workspace, then add:
Use prepare_host.py with an existing verified client, or bootstrap_host.py with an additional new client directory. Keep the original library payload, selected tool/profile bindings and required runtimes available. For configured library requests, explicitly select the relevant families. Acquisition-only readback can use --families '' without restoring acquisition tools. Do not carry live acquisition, Compose execution or prior-configuration options into this readback command. Optional source readers have their own archived contracts and must be enabled separately. The new output/workspace/client must be disjoint from old workspaces, manifests, acquired staging paths and recipe data. Read the exact archive’s selected/contracts/HOST-PREPARATION.md for mixed-request bindings. The preparer checks each selected request; it does not search directories for a plausible installation or migrate to another library version. Start the new generated MCP connection. Use setup_jobs/setup_job to inspect the selected requests, then call setup_reconcile for each relevant job. Imported jobs remain unconfirmed with historical_outcome_unknown. A separate matched observation confirms the checked retained files, not the original operation’s result, account access or tool availability. Readback restores no installation, retry, stop or business-task authority. To use an intact installed connector, the host must validate and authorize its existing launch and attach it through the native owner. A fresh history readback alone does not restore connector processes.

​
Reuse saved settings in a new setup

To continue configuring an integration through this library path, create a fresh setup and add:
Choose the exact configured-install request from the saved file, even if it contains only one entry. This is its request ID, not a guessed directory name. The host approves that request’s full contents for the new setup audience. Keep its original workspace available until preparation completes. Only that request is copied into private prior-request.json; after successful preparation, this new setup no longer needs the original workspace. Choose current families and runtimes explicitly. Reuse prior configuration or setup_successor proposes saved values against current help. Review unapplied settings, select operations again and validate/install into a new directory. The old selection cannot enable a family, authenticate an account, change a running connection or grant new use. Acquisition, selected-package and recipe-export requests cannot seed this configured-library shortcut. An exact raw configured request remains supported through --prior-request and --prior-request-sha256. Do not combine either prior-selection form with readback recovery.

​
Restore software or update a connection

An installed environment and its absolute launch paths must stay in place. To restore software or adopt a new release, install the selected bytes into a new directory, review the current configuration and let the host switch its binding. Retain the old environment until the host has handled its dependants. This does not restore credentials, migrate provider data or replay a write. For interrupted preparation and ambiguous outputs, use troubleshooting.

​
Service connections

For an application management view, compose the following information and controls with your existing account and policy services. A connection view should show the observed account, provider credential scope, effective operations/resources, assigned agent or task, grant lifetime and current diagnostics. A broad provider token can support a narrower local grant only where actual operation enforcement provides that restriction.
1

Inspect the selected account

Match provider-observed identity with the intended host account. A friendly label is not an account verification receipt.
2

Inspect effective use

Review which agent or task can perform which operations on which resources. Keeping the account connected does not extend an expired task grant.
3

Choose the intended control

Reconnect repairs authentication. Change account creates an explicit selection change. Revoke access changes a grant. Disconnect affects the connection and any dependants named by the host.

​
Agent frontends

Frontend management shows the entrypoint, admitted people/groups, target agent, conversation policy, output visibility, service authority and exposure lifetime. Include whether the interface supports text, media, streaming, approvals or interruption and how unsupported interactions are handed off. Pause blocks new admission according to the recorded policy. The host separately determines what happens to admitted work, observation subscriptions, active live sessions and queued replies. Remove ends the frontend binding and preserves unrelated agent state. Before removing the frontend carrying the management conversation, provide another authenticated management route or obtain the specific authorization needed to lose that route.

​
Package pages

Package installation and export activation are different. A package page should link its service connections, frontends and supporting components so that a shared dependency is visible before removal or upgrade. An upgrade selects exact new bytes and configuration, checks migrations and compatibility, and reconciles affected bindings. Retaining the same tool name does not make a changed implementation the same authorized selection.

​
UI outcome handling

The current UI submits the exact displayed owner control once per offered occurrence. A rejected callback means the acknowledgment is uncertain. It does not prove that the owner rejected or rolled back the action. Refresh from the owner before offering recovery; do not turn a rendering error into a new setup attempt. See UI integration, configuration and the operations runbook.