Building one workspace for AI work
Connecting EMO governance, a working Paperclip dashboard and a local Hermes tool library into a workflow people can inspect and review.
Body
A useful AI workspace should make a job easy to follow. What was requested? Which information did it use? What was allowed to happen? Where is the result, and who decides whether it is ready?
At Mathura Media Group, we are connecting those questions to a working local pilot. Paperclip holds the project work list. Codex helps prepare and coordinate the work. Hermes provides an installed agent environment connected to a local model. EMO supplies the authorization and execution boundary for the governed workflow.
The project has produced a functioning task connection, a local tool and skill library, and a recorded trial through EMO. Automatic delegation is still being developed. The practical progress is that the work, its permissions and its actual outcome can be inspected together.
A shared place to see the work
Paperclip is an open-source application for coordinating agent work. Our local installation has three project areas covering an MMG catering prototype, Rubare Tutto and EMO's governed office workflow.
The Codex connection can read projects and tasks, create unassigned draft tasks and record progress notes. We have verified the native connection against the running dashboard, including reading existing work and saving a note that can be read back. Earlier bridge tests also checked duplicate-write protection and the project boundary.
This gives a request somewhere durable to live. A person can see the task and its evidence without reconstructing a conversation from memory. The dashboard is being used today; automatic worker execution remains paused.
Local capability with a visible inventory
Hermes is installed, and a read-only check confirmed that the selected local model was available. We have also downloaded an offline library containing official Hermes tool source, bundled skills, optional skills and supporting documentation.
The library includes the source release matching our installation and a separately pinned current official snapshot. Optional dependency downloads resolved for 32 of the installed release's 34 backend groups. The resulting cache contains 201 distinct package files whose hashes and archive integrity were checked. Two messaging integrations still need build work.
This preparation makes the source and many installation dependencies available for future local setup. Downloaded files do not automatically grant a worker permission to use them. Services such as web search and remote APIs still need connectivity and any required service account.
EMO makes the execution boundary concrete
EMO is the system we are developing to connect requests, permissions, selected inputs, execution and evidence. In the governed path, a job needs a defined scope and an expected result. Its outcome must distinguish refusal, failure and a draft awaiting human review.
We tested that path with one explicitly approved internal catering briefing. The request passed EMO authorization, capability resolution and handler selection, then reached the local Hermes worker. The worker requested tools that were unavailable in that test configuration and did not return a complete briefing.
The output checks rejected the incomplete result. The failed state and evidence were saved, and the temporary execution changes were restored. No customer outreach or publication occurred as part of that trial.
That result showed both the working connection to the worker and the next engineering task. We have prepared a revised prompt that supplies the relevant facts directly. It has passed focused offline checks, but it has not yet been validated in a fresh live trial.
A workflow people can review
The intended workflow gives each component a clear job: Paperclip tracks the work; Codex helps prepare it; EMO checks and controls the governed execution; Hermes performs selected work within its configured boundary; a person reviews the result.
We have also prepared a read-only terminal status view. It reports the approval state, execution setting and recorded trial outcome without restarting a worker. Together, the dashboard and status view help distinguish a successful diagnostic from a failed model task. Across the prepared integration changes, 71 focused offline checks have passed; the live briefing outcome remains failed.
For a catering campaign, review includes the offer, menu, delivery terms and audience. For a brand post, it includes the artwork, wording and purchase route. Those decisions remain visible parts of the work rather than assumptions hidden inside a model response.
The next stage
The next milestone is a useful briefing returned through EMO, checked against its selected inputs and presented for human review. The revised prompt needs a newly authorized live trial before we can claim that milestone. Automatic workers remain off in this pilot.
What we have built so far is a practical foundation: a shared work list, a verified task connection, locally stored capabilities and an execution trail that records what actually happened. MMG is using that foundation to develop AI workflows whose progress people can see and whose next steps they can deliberately choose.