Sequential Thinking MCP Server

MCP Sequential Thinking

A local MCP server for working notes you can resume: evidence, decisions, risks and next actions for a difficult task, stored durably in SQLite and readable after any restart. It records what the caller writes; it does not generate, judge or improve reasoning, and it never asks for hidden chain-of-thought.

Installation

Published on PyPI and in the MCP Registry as io.github.arben-adm/mcp-sequential-thinking. With uv installed, uvx mcp-sequential-thinking runs it without a manual install (pin with uvx mcp-sequential-thinking@0.7.0). MCP_STORAGE_DIR is optional.

Claude Code

claude mcp add working-notes -- uvx mcp-sequential-thinking                   # this project only
claude mcp add --scope user working-notes -- uvx mcp-sequential-thinking      # all projects
claude mcp add --scope project working-notes -- uvx mcp-sequential-thinking   # shared via .mcp.json

Gemini CLI

gemini mcp add working-notes uvx mcp-sequential-thinking   # writes .gemini/settings.json

Codex CLI

codex mcp add working-notes -- uvx mcp-sequential-thinking

Claude Desktop (Settings → Developer → Edit Config, claude_desktop_config.json)

{"mcpServers":{"working-notes":{
  "command":"uvx",
  "args":["mcp-sequential-thinking"],
  "env":{"MCP_STORAGE_DIR":"/absolute/path/to/notes"}}}}

Restart Claude Desktop afterwards.

VS Code (.vscode/mcp.json)

{"servers":{"working-notes":{"type":"stdio","command":"uvx","args":["mcp-sequential-thinking"]}}}

If a GUI client cannot find uvx, use its absolute path (which uvx).

From source

python -m venv .venv && . .venv/bin/activate   # Windows: .venv\Scripts\Activate.ps1
python -m pip install -e .
python scripts/stdio_smoke.py                  # restart demo: write, restart, read back

Then point the client at /absolute/path/to/.venv/bin/mcp-sequential-thinking (Windows: .venv\Scripts\mcp-sequential-thinking.exe) instead of uvx.

Quick start

A typical session:

{"tool":"create_session","arguments":{"title":"Choose a local store","request_id":"create-1"}}
{"tool":"add_step","arguments":{"session_id":"<id>","content":"Two writers need unique positions.","kind":"evidence","request_id":"step-1"}}
{"tool":"add_step","arguments":{"session_id":"<id>","content":"Use SQLite.","kind":"decision","request_id":"step-2"}}
{"tool":"read_session","arguments":{"session_id":"<id>"}}

Lost the ID after a restart? list_sessions finds the session by title.

Tools

Links between notes. parent_step_id records provenance, supersedes_step_id replaces an earlier note (resume hides it, full history keeps it), and branch_from_step_id starts a named branch. All references stay inside one session.

Retries. Every write takes a request_id. Repeating a call with the same ID and input returns the original result; the same ID with different input returns IDEMPOTENCY_CONFLICT. finalize_session and delete_session also need expected_version.

Pagination and limits. Reads default to 20 entries and 12,000 characters (limit ≤ 100, max_chars 1,000–50,000). Follow cursor until it is null, and keep the same view while paging. truncated reports shortened content. A note too large for one response comes back as its first chunk; continue with step_id and content_offset until next_content_offset is null.

Errors set isError=true with a stable code and a short hint, e.g. INVALID_INPUT (with Fields: …), UNKNOWN_SESSION, INVALID_REFERENCE, CONFLICT (with current_version), SESSION_FINALIZED, IDEMPOTENCY_CONFLICT, STORAGE_BUSY (retry with the same request_id), PATH_OUTSIDE_EXPORTS, RESPONSE_LIMIT.

Sources attached to notes are caller-supplied and never fetched or verified.

Legacy tools (deprecated)

process_thought, generate_summary and clear_history, plus export_session / import_session without a session_id, work on a single shared legacy session. They remain for existing clients; every result carries a deprecation_notice. Use the session API for new work.

• Stages: Problem Definition, Research, Analysis, Synthesis, Conclusion (case-insensitive). Order warnings are advisory unless --strict-stages is set.
• main_line_progress counts mainline notes only, against your total_thoughts; it is not task completion.
• Legacy import replaces the legacy history.
• read_session(session_id="legacy") reads it; a legacy branch X appears as legacy-X there.

Storage and operation

mcp-sequential-thinking              # stdio server
mcp-sequential-thinking --health     # diagnostic check
mcp-sequential-thinking --ephemeral  # in memory only, lost on exit

• Data lives in ~/.mcp_sequential_thinking (or MCP_STORAGE_DIR) as plaintext SQLite (WAL, synchronous FULL, 5 s busy timeout). Use a local disk; NFS/SMB and multiple hosts are not supported.
• Several stdio servers may share one store. Local stdio trusts the OS user; a session ID is not authorization. HTTP/SSE are experimental.
• Deletion is not secure erasure. There is no automatic retention.
• On first start, an existing 0.6 JSONL history migrates into SQLite in one transaction, keeping IDs and content; a backup of the source stays in place. Stop old 0.6 servers first. Downgrading after new writes is not supported.

Backup and restore (all sessions, including retry keys):

python -m mcp_sequential_thinking.backup create  /path/to/storage /path/to/snapshot.sqlite3
python -m mcp_sequential_thinking.backup restore /path/to/NEW-storage /path/to/snapshot.sqlite3

Restore requires a new directory. Never copy a live database file without its WAL.

Tested with

Development

uv run --extra dev ruff check .
uv run --extra dev ruff format --check .
uv run --extra dev mypy --strict mcp_sequential_thinking
uv run --extra dev pytest --cov=mcp_sequential_thinking --cov-fail-under=85
rm -rf dist && uv run --extra dev --with build --with twine --with jsonschema python scripts/check_artifacts.py

The artifact check builds and installs the wheel in isolation and runs a real restart round trip. Releases publish that tested artifact via Trusted Publishing once a GitHub release is created; the version lives in mcp_sequential_thinking/_version.py. See CHANGELOG.md for details.