Features / MCP Server

Your AI remembers

The MCP server lets AI agents search your past sessions and read project context during conversations. One command to install.

mcp install
$ sfs mcp install --for claude-code
MCP server configured for Claude Code.
Restart Claude Code to activate.
68 tools available to your AI agent:
Session tools (7):
• search_sessions — Search past sessions
• get_session_context — Full conversation from a session
• list_recent_sessions — Browse recent work
• find_related_sessions — Find by file or error
• get_session_summary — Structured session summary
• get_audit_report — Trust audit report
• get_session_provenance — Which rules instructed a session
Knowledge read (9):
• get_project_context — Full wiki: overview + pages + concepts
• get_context_section — One section of the context doc
• get_wiki_page — One wiki page in full
• get_wiki_page_history — Revision history with persona / ticket attribution
• search_project_knowledge — Search knowledge entries
• list_knowledge_entries — Filter by type / class / freshness
• get_knowledge_entry — One entry's full record
• get_knowledge_health — Pending / compiled counts
• ask_project — Q&A with sources_cited
Knowledge write (5):
• add_knowledge — Save a discovery / decision / pattern
• update_wiki_page — Create or update a wiki page
• list_wiki_pages — Browse wiki structure
• compile_knowledge_base — Trigger a compile pass
• dismiss_knowledge_entry — Dismiss a stale or wrong entry (audited)
Rules (2):
• get_rules — Canonical project rules
• get_compiled_rules — Compiled rules per tool
Personas (5) — Pro+:
• list_personas — Project personas
• get_persona — One persona's full content
• create_persona — Define a new project persona
• assume_persona — Tag captured sessions with this persona
• forget_persona — Clear the active persona bundle
Tickets (9) — Team+:
• list_tickets — Filter by status / assignee
• get_ticket — Ticket detail + dependencies
• create_ticket — Open new ticket (acceptance criteria required for agents)
• start_ticket — Atomic claim + compiled context + bundle write
• complete_ticket — Send to review, clear owned bundle
• resolve_ticket — Mark done from review
• assign_persona — Set the persona on a ticket
• escalate_ticket — Audit comment + raise priority
• add_ticket_comment — Slack-like thread
Work queues (6):
• create_work_queue — Aim an agent at a queue of tickets
• get_work_queue — Queue config + progress
• list_work_queues — Browse a project's queues
• set_work_queue_status — Pause / resume / archive
• run_work_queue_step — Heartbeat: one bounded directive per wake
• complete_work_queue_step — Durably advance the queue

What the MCP server gives you

One-command install

Works with Claude Code, Cursor, and Copilot. Adds SessionFS to your tool's MCP config file automatically — no manual JSON editing.

Project context

Share architecture decisions, coding conventions, and API contracts with every AI agent on your team. One document per repo, always current.

Local + Remote

Local MCP server on stdio for desktop tools — no network required. Remote HTTP/SSE server for cloud-connected AI tools and team sharing.

How agents use session memory

When you start a new session, your AI agent can automatically search for related past sessions and load their context — so it already knows what you tried last time.

The find_related_sessions tool maps file paths and error strings to previous sessions. If you're debugging the same file you touched last week, the agent knows.

Read the MCP docs
agent using MCP
Agent calls MCP tool
find_related_sessions ("tokens.py")
SessionFS returns
ses_abc — Debug auth middleware (2 days ago)
ses_d2e — JWT token refresh fix (5 days ago)
Agent says
"I found 2 related sessions. The last time you worked on tokens.py, you extended the expiry window and added redis for caching..."