One MCP server instead of eight.

200 tool definitions in every context window. Rostr replaces them with four and a playbook that learns from your runs.

$npm install -g rostr-mcp
$rostr init

Add to .cursor/mcp.json, reload.

Before / after

Without Rostr

MCP servers loaded: 8
Tool schemas in context: ~200
Tokens burned on tools: ~87k
Memory across sessions: none

With Rostr

MCP servers loaded: 1
Tool schemas in context: 4
Tokens burned on tools: ~4k
Memory across sessions: .mdc playbook

How it works

Discover

Reads your .cursor/mcp.jsonand plugin directories. Builds a roster of what's actually connected right now.

Plan

suggest_plan looks at the goal, your servers, and your run history. Returns a sequence with warnings where things have broken before.

Execute

The agent still calls Neon, Stripe, GitHub directly. Rostr doesn't proxy HTTP. It routes attention, not packets.

Learn

log_run records outcomes. Patterns get extracted. .cursor/rules/rostr.mdc gets rewritten so the next turn starts with context.

Tools

name	when	what
list_roster	Start of a task	Connected servers, saved workflows, recent stacks
suggest_plan	Before multi-step work	Optimal sequence with failure warnings from history
log_run	After a run	Persists outcome, extracts patterns, rewrites .mdc
recall_playbook	Unfamiliar stack combo	Known patterns and success rates for that set of servers

Pattern learning

Every logged run gets fed through rule-based extraction. Failure rates, ordering dependencies, timing quirks. No model calls, no cloud. SQLite and a rules file that Cursor already reads.

rostr patterns

run_sql fails after fast runs — space steps
create_branch → run_sql works in that order
Neon stack: 43% success over 7 runs

$npm install -g rostr-mcp && rostr init

Local only. MIT licensed. No account, no API keys, no telemetry.