A few habits make the DevStride MCP Server faster, more accurate, and easier to trust.
Lead with questions that only read — "what's assigned to me?", "summarize my priorities", "what's on the Platform board?". They are fast and safe, and they show you how the assistant interprets your requests before you let it change anything.
Ask "What tools are available from the DevStride MCP, and what can you do with them?" — the built-in devstride_help tool returns a current catalogue of prompts grouped by job. The Tool Reference lists every tool.
Two behaviours worth knowing:
invoke_api_operation and includes a misspelled or misplaced key, the call is refused up front and the reply names every key that would have been ignored. The curated tools warn after the call instead when they drop an unknown top-level argument.The connection has your permissions, so the assistant can change real work. For anything that writes — estimates, dates, bulk edits, a new roadmap — end the prompt with "…propose first and wait for my approval before changing anything." You see the plan, then say go.
Prefer reversible tools. archive_item, archive_board, and archive_workstream can be undone; delete_item, delete_board, delete_workstream, and delete_time_entry cannot. If the assistant proposes a permanent delete, ask whether archiving would do.
You do not need IDs. The assistant resolves "move this to the Sprint 2 board", "assign it to Priya", "tag it Blocked". Spell names the way they appear in DevStride; when a name is ambiguous the assistant is told the candidates and will ask you which one.
Item numbers are the exception worth knowing: an I number is a work item and an F number is a workstream, each with its own tools. Sending a workstream number to an item tool is refused with a message naming the right tool.
/exit, start claude again, then /mcp to confirm DevStride is connected./mcp route itself is throttled at the edge to bound a runaway client. Over the limit, the MCP tells the assistant how long to wait and carries on once the window resets — see API keys.bulk_update_items applies up to 400 work units per call; larger ones are refused before anything changes, and batches too slow to finish inside one request are queued as a job the assistant polls with get_job. A queued batch is not done until the job reports it is.create_workstream not create_folder, create_status not create_lane. The old folder- and lane-named tools were removed on October 10, 2026, and rank_item no longer accepts the scope lane (use status). Update any saved prompts and scripts that still name them — see API keys.Doing Real Work with the DevStride MCP Server
What you can ask the DevStride MCP Server, grouped by the job you are doing, with a worked example from backlog to roadmap and prompts by role.
Tool Reference
Every tool the DevStride MCP Server exposes, grouped by job, with what each one does and which groups are hidden from the core toolset.