Continuity
A new contributor can read the task and recent evidence instead of relying on a copied conversation or an old summary.
SHARED MCP + OPENPROJECT
Many contributors.
One managed workflow.
Shared context helps people and agents collaborate. Shared governance makes that collaboration accountable. Our OpenProject work provides a practical example of the approach.
THE CONNECTION LAYER
Common tools.
Individual boundaries.
Model Context Protocol (MCP) gives AI applications a common interface to tools and context. Multiple agents can connect to a shared service while retaining separately scoped identities and permissions.
Human project ownerResearch agentEngineering agentReview agent
Scoped MCP interfaces Identity · permissions · attribution
Managed project record
Goal → proposal → decision → evidence → handoff
Conceptual architecture. MCP supplies the interface; service policy and workflow controls supply authorisation and coordination.
THE OPENPROJECT EXAMPLE
The work item
outlives the chat.
In our development work, a scoped project connector lets assistants read the same work items and save attributed proposals and handoffs. People can respond in the project record, and another connected assistant can pick up that context.
The useful pattern is a shared source of work: the objective, latest answers, decisions and evidence stay attached to the task. A saved handoff and a launched worker are distinct states.
- 01
Define the work
The owner records the goal, constraints and acceptance criteria.
- 02
Read before acting
An agent reads the current task and recent answers through the scoped connector.
- 03
Make a durable handoff
Findings, uncertainties and next steps are saved as attributed project notes.
- 04
Decide and verify
Human authority governs the next action. Results are checked and linked back to the work.
WHAT THE SHARED PATTERN MAKES POSSIBLE
Less lost context.
More accountable progress.
Clear attribution
Agent proposals, human answers and verified outcomes have distinct roles. A suggestion is never silently promoted into a decision.
Controlled coordination
Task records guide collaboration. Explicit grants, budgets and execution checks control what can actually run.
Further reading: MCP architecture · OpenProject work packages. This example describes our development approach and is not an OpenProject endorsement.
A SHARED INTERFACE, CLEAR RESPONSIBILITIES
Give each contributor
the access its role needs.
A shared MCP service can make approved context and tools available to different agent clients. Useful collaboration also depends on identities, permissions and a managed record of what each contributor did.
Design tools around the work.
A project integration might expose separate operations to read a work item, retrieve recent notes and save an attributed proposal. Each operation has a clear purpose and a checked input contract.
Give each contributor a scoped identity where the integration supports it. Use service-side checks to enforce project boundaries and permitted operations, while keeping credentials outside prompts and shared task records.
Coordinate through durable state.
MCP provides access to context and tools. The project workflow records ownership, decisions and progress; execution controls determine what can run. Design those parts together so a tool connection does useful work within an agreed boundary.
Before contributing, read the current task and recent answers. Before applying a change, recheck that the approved scope and underlying evidence are still current. Another contributor may have moved the work forward.
A HANDOFF THAT ANOTHER CONTRIBUTOR CAN USE
Preserve the reasoning.
Make the next step clear.
01 / OUTCOME
What was learned?
State the result, its source evidence and any uncertainty. Separate observed facts from proposed explanations.
02 / DECISIONS
What did the owner decide?
Link the human decision and its scope. Keep agent suggestions and permission to execute distinguishable.
03 / VERIFICATION
What was actually checked?
Record the checks performed and what they establish. Identify acceptance steps that remain outstanding.
04 / NEXT STEP
Who can move the work forward?
Name the next action, responsible contributor and any blocker. A saved proposal, a queued job and a completed result are different states.
In a managed shared workflow, an agent can prepare the investigation, another can review it and the owner can decide the next action. OpenProject carries the work record across those handoffs, helping the team continue without reconstructing an entire chat.
LET’S FIND YOUR FIRST WORKFLOW