Task context and handoffs over MCP
Agents ask for a task's context before they start, and leave a handoff when they stop.
Give coding agents the same record your engineers use
Give coding agents the spec, plan and open questions your engineers work from, plus the task's context with its labels. Agents read the record over MCP, add attributed, agent-reported work beside the plan, and leave a handoff for the next session.
Plans: The MCP server is on every plan. One design doc on Free; unlimited design docs and Ask AI on Team and above. GitHub evidence on Business and Enterprise.
- 1. Asks for context
- Task: shorten the tender window for API carriers. Current: DD-0021, approved by a person; ADR-0003. Open question: Q1, assigned to Maya Chen.
- 2. Takes a task
- Channel tag in Carrier Gateway, from the implementation plan.
- 3. Reports
- Marks the task done. Recorded as Agent-reported; no recorded change for Carrier Gateway yet.
- 4. Leaves a handoffEarly access
- E-4 by Claude Code, run by Priya Raman: Done, Blocked on Q1, Next: confirm the Carrier Gateway change.
Start agents with the agreed design
Read the spec, plan and open questions over MCP.
See what the agent reported
Keep attributed recorded changes beside the implementation plan.
Keep decisions with people
Agents raise open questions; people answer them and approve variations.
Start with the design and the open questions
Ask an agent to read the records linked to a change before proposing how to build it. The MCP server gives it the architecture people use to review that work.
- Docs and diagrams
- Catalog objects, dependencies and impact
- Design doc specs, plans and open questions
- Decision records and recorded changes
Put the work back into the record
An agent can create and edit diagrams as YAML, write design doc variations with a brief summarizing the full text, and record architecture changes.
- Import a Claude Code plan, Spec Kit plan.md, Kiro tasks.md or markdown plan file
- Link pull requests to the work
- Submit evidence labeled as the agent's report
Plans: Recorded changes on every plan. Linked GitHub evidence on Business and Enterprise.
Acceptance criteria on every plan, up to 10 per design doc on Free. Early access
People keep the review decisions
Agents never approve variations or open amendments. Open questions go to a person.
Attribution
Every write records the person, the client and an optional session link.
Review
An agent's report is recorded evidence. It is separate from a verified result.
Linear drafts
A person confirms draft issues before they become execution work in Linear.
v4 · Approved by the Platform team
ApprovedAmendment · Refund window changed to 30 days · back in review
In review- v3 to v4 · 2 sections changed · agent write from Claude Code
Linear on Business and Enterprise. Early access
Archdocs plugin skills
Skills for design and close-out
The Archdocs plugin for Claude Code, Codex and Cursor includes skills plus /archdocs:design-doc and /archdocs:close-out.
Get the Archdocs plugin on GitHubAsk about the architecture, with citations
Ask AI answers questions about your architecture with citations and can generate diagrams. Use it when you want to explore the record yourself.
- Follow citations back to the source docs
- Generate a diagram from a description
Plans: Ask AI on Team and above, using AI credits.
Connect your client
Use https://mcp.archdocs.io/mcp. Sign in with OAuth or a personal access token. Claude Code, Codex, Cursor and VS Code can connect to the hosted server.
Screenshots are from a sample workspace. Recreated views are labeled.