Skip to content

MCP.md

Latest commit

 

History

History
116 lines (85 loc) · 4.38 KB

MCP.md

File metadata and controls

116 lines (85 loc) · 4.38 KB
Raw

MCP — Thread Memory

Paper_Rec ships a Thread Memory MCP (packages/thread-mcp) so Cursor / Claude Desktop / other MCP clients can read and update Cognitive Threads without owning the whole Wiki UI.

Why not another search MCP?

article-mcp already covers multi-source literature search. Paper_Rec MCP focuses on what only we have: hypothesis, claims, Claim–Evidence Map, evidence gaps, ledger gates, lit↔exp membership, and Watch/Delta.

Recommended composition:

  1. article-mcp (or Skill /query_*) → candidate papers
  2. thread_query_hint / thread_search_context / thread_score_papers → Thread-conditioned judgment
  3. thread_link_* + thread_add_evidence after human/agent gate

Session pattern (gptr-mcp inspired)

Prefer a stable thread_id / research_id for the turn:

  1. thread_show / create → bind session
  2. quick path: thread_search_context + score (shallow)
  3. deep path: multi-path /query_* + evidence adds + claim-ledger
  4. Always expose sources (paper_paths / cite keys) before claiming context is grounded

Deferred gather→write (disk-backed):

python -m wiki_bridge.cli research-session --wiki-root ../.. --action create --topic "..." --sources-json papers.json
# later
python -m wiki_bridge.cli research-session --wiki-root ../.. --action sources --research-id <id>
python -m wiki_bridge.cli research-session --wiki-root ../.. --action write-report --research-id <id>

Store: content/_meta/research_sessions.json (TTL ~7d). Do not invent a second research MCP; chain Thread tools with Skill retrieval.

Quick config (2.16+)

推荐：安装后先 dry-run，再 --apply 写入。

# Windows
powershell -ExecutionPolicy Bypass -File scripts/configure-mcp.ps1
powershell -ExecutionPolicy Bypass -File scripts/configure-mcp.ps1 -Apply
# macOS / Linux
chmod +x scripts/configure-mcp.sh
./scripts/configure-mcp.sh
./scripts/configure-mcp.sh --apply
# after pip install -e packages/thread-mcp:
paper-rec-configure --apply
  • 默认目标：docs/mcp.example.json + 项目 .cursor/mcp.json
  • 全局 Cursor / Claude Desktop：--target cursor-user / --target claude-desktop（建议加 --force，会写 .bak）
  • VS Code Continue：见下方提示，手动合并同一 mcpServers 块

No PYTHONPATH required. Set PAPER_REC_ROOT to the workspace root; the server auto-adds packages/wiki-bridge.

{
  "mcpServers": {
    "paper-rec-threads": {
      "command": "python",
      "args": ["-m", "thread_mcp.server"],
      "env": {
        "PAPER_REC_ROOT": "/path/to/Paper_Rec_Skill"
      }
    }
  }
}

Install once:

pip install -e packages/wiki-bridge -e packages/thread-mcp
pip install "mcp>=1.0"

See packages/thread-mcp/README.md.

Var Meaning
PAPER_REC_ROOT Workspace root containing content/threads/

Tools summary

Tool Role
thread_list / thread_get List / full state + evidences
thread_search_context / thread_query_hint Context + query hints for external search
thread_score_papers / prerank_papers Score vs thread / BM25 pre-rank
thread_link_paper / thread_link_exp Membership
thread_add_evidence / evidence_coverage Claim–Evidence Map + confidence UX + CEBM-lite
thread_graph / bibtex_export / related_work / section_outline / paper_draft Graph + writing frames
citation_expand 1-hop refs (S2/Crossref)
thread_delta / thread_claim_* Watch + claim gates
wiki_list_papers Local wiki cards
exp_list / exp_get_metrics Local experiment metrics

Compose with search MCPs

Paper_Rec MCP = memory. For multi-source search / OA PDF download chains, also install:

  • scholar-mcp (uvx scholar-mcp) — RRF search + PDF tools
  • or PaperSeek (paperseek-mcp) — iterative literature agent

Then: search MCP → candidates → rrf_fuse / prerank_papers / thread_score_papers → thread_link_* / pdf_fetch / thread_add_evidence.

Local OA fetch (no Sci-Hub): Wiki「获取全文」or wiki_bridge pdf-fetch.

Related