threadctx-mcp

Shared memory MCP server for AI coding agents. Works identically with Claude Code, Cursor, and any MCP client — same package, same config shape, no per-client integration work. On first start it also drops a "check team memory" instruction into whichever agents' rule files your repo uses (AGENTS.md, CLAUDE.md, Copilot, Windsurf, Cline, Gemini) so the memory actually gets read, not just exposed.

This repo runs on itself: its own AGENTS.md and CLAUDE.md were set up with npx threadctx-mcp init, the same command you'd run — check the git history if you want to see it happen.

Modes

• Local (default, free, no signup): memory stored as a plain JSON file at ~/.threadctx/local.json — zero native dependencies, so npx threadctx-mcp installs instantly on any machine with Node 18+ (no compiler, no node-gyp step). No network calls except to whichever LLM provider your agent already uses. Matching is keyword-based, scoped to the current repo (detected via git remote). Run npx threadctx-mcp list any time to see exactly what your agents have stored.
• Cloud (paid Team tier+): memory shared across everyone on the repo, with real semantic search. Requires an API key from threadctx.dev (or your own self-hosted deployment — see ../cloud/README.md).

Quick start

# Local mode — nothing to configure. Also auto-adds the "check team memory"
# instruction to your agents' rule files (AGENTS.md, CLAUDE.md, and any
# detected tool-specific files) on first start.
npx threadctx-mcp

# Set a repo up for your whole team (run once, commit the results)
npx threadctx-mcp init

# Cloud mode — shared team memory via threadctx.dev
npx threadctx-mcp init --mode=cloud --api-key=tctx_xxx

# Joining a repo a teammate already set up? One command:
npx threadctx-mcp join

# See what your agents have written to this machine
npx threadctx-mcp list            # this repo
npx threadctx-mcp list --all      # every repo

init writes three committable, secret-free files: .threadctx.json (just { "mode": ... }), .mcp.json (Claude Code project config), and .cursor/mcp.json (Cursor project config). Commit all three — every teammate who then opens the repo in Claude Code or Cursor is prompted to enable threadctx automatically, with nothing to install or configure. Teammates on other MCP clients run npx threadctx-mcp join, which sets up their machine the same way and prints the config block for their client.

Your API key never appears in any committed file. It is read from the THREADCTX_API_KEY environment variable at runtime (export it in your shell profile), so secrets stay out of version control by construction.

threadctx also adds a small, clearly-marked instruction to your project rules telling the agent to call memory_query before a task and memory_write after. It writes the two universal files every time — AGENTS.md (the cross-tool standard read by Copilot, Cursor, Windsurf, Zed, Codex, Aider, and ~24 others) and CLAUDE.md (Claude Code's richer native format) — plus .cursor/rules/threadctx.mdc. It then adds a tool-specific file only when that tool's footprint is detected in the repo, so it never litters your project with rule files for tools you don't use:

Shared files (AGENTS.md, CLAUDE.md, Copilot, Gemini) get a marker-fenced block spliced in, preserving your own content around it; dedicated files are owned in full. This happens automatically the first time the server starts in a project — you don't need to run init for it. Running init just triggers it explicitly and prints the result; either way it's idempotent (safe to re-run, never duplicates). Opt out entirely with THREADCTX_NO_AUTO_RULES=1, or per-init-call with --no-rules.

The same instruction is also sent as part of the MCP initialize handshake itself (the protocol's instructions field), so it reaches the model even before any rules file exists, and for clients that don't read project-rules files at all. The file-based rules are belt-and-suspenders on top of that, since not every MCP client is guaranteed to surface instructions prominently.

Claude Code setup

Add to your Claude Code MCP config (claude mcp add or edit ~/.claude/mcp.json directly):

{
  "mcpServers": {
    "threadctx": {
      "command": "npx",
      "args": ["-y", "threadctx-mcp"],
      "env": {
        "THREADCTX_MODE": "cloud",
        "THREADCTX_API_KEY": "tctx_xxx"
      }
    }
  }
}

Cursor setup

Add the same block to .cursor/mcp.json in your project root (or via Cursor Settings → Tools & MCP):

{
  "mcpServers": {
    "threadctx": {
      "command": "npx",
      "args": ["-y", "threadctx-mcp"],
      "env": {
        "THREADCTX_MODE": "cloud",
        "THREADCTX_API_KEY": "tctx_xxx"
      }
    }
  }
}

That's it — the same package and config work in both clients because MCP is a portable, open protocol.

Passive capture — turn git history into memory

Memory shouldn't depend on an agent remembering to call memory_write. threadctx capture reads the commits landed since its last run, uses your own LLM provider key to distill the genuinely reusable decisions and gotchas (skipping trivial commits), dedups them against what's already stored, and writes the survivors. It's tool-agnostic — it doesn't matter whether the work happened in Claude Code, Cursor, Copilot, or a plain editor.

# Off by default because it calls an LLM (billed to your provider). Enable it:
export THREADCTX_CAPTURE_ENABLED=1
export ANTHROPIC_API_KEY=sk-...     # or OPENAI_API_KEY

npx threadctx-mcp capture --dry-run     # preview what it would store
npx threadctx-mcp capture               # store them (incremental since last run)
npx threadctx-mcp capture --since=v1.2.0 --diffs   # a range, with patches

# Scaffold a GitHub Action that captures every merged PR automatically:
npx threadctx-mcp capture --print-workflow > .github/workflows/threadctx-capture.yml

Capture calls your LLM provider directly — nothing is routed through threadctx's servers, so local mode keeps its "no network call beyond your own LLM provider" promise. It is off unless THREADCTX_CAPTURE_ENABLED=1 (or a one-off --force), so it can never run up token cost as a side effect.

Environment variables

Local development

npm install
npm run dev     # runs the server via tsx, watches for changes
npm run build   # compiles to dist/ for publishing

How the tools work

• memory_write(content, tags?) — the agent calls this after resolving a non-obvious bug, making an architectural decision, or learning something worth remembering.
• memory_query(task_description, max_results?) — the agent calls this before starting risky or repeated work. Results are returned with a consistent attribution footer (· via threadctx — shared team memory (N hits)) so the same string is recognizable whether you're reading Claude Code's terminal output or Cursor's agent panel.

Tool descriptions are written to bias the model toward calling memory_query proactively, and — as of 0.3.0 — the server reinforces this two more ways with zero setup required: the MCP initialize response carries the same instruction to every connecting client, and CLAUDE.md / .cursor/rules/threadctx.mdc get it auto-injected on first start. MCP tools are still fundamentally pull-based (no mechanism can force a tool call), but these three layers together are the strongest guarantee we can build.

CLI subcommands