MCP server: pass structured context between AI coding agent sessions. No more starting from scratch.
Three tools with complete input schemas and descriptions. Naming follows verb_noun pattern (baton_pass, baton_receive, list_batons). Descriptions are adequate (100-150 chars) but lack actionable context for LLM selection. Parameters are typed and mostly described, but lack constraints (enums, ranges, patterns). Output schema is undocumented, LLMs cannot predict response structure. Error handling is minimal; no recovery guidance. No tool annotations (readOnlyHint, destructiveHint). baton_pass modifies state but description doesn't explicitly warn of side effects.
Save a structured context snapshot so the NEXT agent session picks up exactly where you left off. Call at the END of a work session.
Retrieve the latest context snapshot for a tag. Call at the START of a new session to get full context from the previous agent.
List recent handoff snapshots for a project.
Output schemas undocumented. LLMs cannot predict response structure (fields, types, pagination). baton_pass returns an object with id, tag, project, agent_name, created_at; baton_receive returns an array of objects with summary, completed, pending, etc., but this is inferred from code, not declared in tool definition.
baton_pass description does not state it modifies state (creates a record). LLMs need explicit 'This tool creates/saves/persists' language to understand side effects and retry safety. Current: 'Save a structured context snapshot', passive voice obscures the mutation.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). baton_pass is destructive (overwrites prior snapshots for same tag); baton_receive is read-only; list_batons is read-only. Annotations enable agents to reason about safety and retry logic.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Parameters lack constraints. 'tag' accepts any string (no length, pattern, or format). 'limit' is unbounded (no min/max). 'project' has no enum or validation. Free-form strings invite hallucinated values; LLMs cannot self-correct without explicit constraints.
Error handling is minimal. baton_receive returns a plain text message if no baton found ('No baton found for tag=...'). No recovery guidance, no list of available tags, no actionable next step. Per pattern:recovery-guide, errors should guide the LLM toward resolution.