Threadwell also exposes get_knowledge for extracted constraints, preferences and gotchas, and get_project_context for a bounded context draft with sources and dates. Both use existing caller permissions and require human review before extracted text becomes standing instructions.
Use it from your assistant (MCP)
TokenTracker exposes your own knowledge to your AI assistant over the Model Context Protocol — so Claude Desktop, Cursor, etc. can recall how you work and what you've decided.
Add the MCP endpoint:
https://tokentracker-console.vercel.app/api/mcp
Read-only, tenant-scoped, RLS-enforced — your assistant only ever sees what you're allowed to see. Tools include:
memory_health— Measured capture, extraction, belief, link and truncation coverage.list_projects— Your projects — resolved, activity-tagged & linked across every tool.list_sessions— Recent sessions. Use the exact project_key from list_projects to filter before the session limit.get_session— A full redacted transcript — turns + artifacts. Secrets never present.activity_timeline— Daily cross-tool activity — sessions, turns, tokens over time.search_my_turns— Hybrid lexical + semantic recall across everything you've ever asked an AI.working_style— Your distilled working-style fingerprint — how you steer AI.get_decisions— Verbatim decisions. Use project_key before query and limit so older project history is included.get_trajectory— A session's build-story — the specify→build→verify→accept arc.export_profile— A ready-to-paste SKILL.md of how you work — drop into any assistant.get_beliefs— What this workspace currently believes, and what reversed it.get_gotchas— Traps already hit and paid for — so they aren't hit twice.get_system— One machine's passport: identity, freshness, capture coverage, 30-day usage.find_similar_sessions— Sessions that resemble this one — prior art for the problem in hand.decision_durability— Which decisions held over time and which quietly got reversed.tool_switching— Where work moves between AI tools, and what triggers the switch.get_skill_candidates— Repeated patterns worth promoting into a reusable skill.link_candidates— Sessions that probably belong to the same project, ranked by evidence.
See Privacy & what ships for the scoping guarantees.