Session-Buddy

Session-Buddy is a session-lifecycle and memory MCP server for Claude Code and other MCP clients. It manages session startup, checkpoints, cleanup, searchable reflections, cross-project context, and quality signals through a local DuckDB-backed service.

Quick Links

• Capabilities

• MCP Surface

• Configuration

• Automatic Session Management

• Integration with Crackerjack

• Quality Checks

• Documentation

Related MCP server: Central Brain

Quality Checks

Crackerjack is the canonical quality gate for repository changes. Use the focused checks while iterating and the full gate before handoff:

crackerjack lint
crackerjack typecheck
crackerjack security
crackerjack run --run-tests

Capabilities

Session lifecycle

• Initialize, checkpoint, inspect, and end sessions through MCP tools.

• Detect Git repositories and perform lifecycle setup and cleanup automatically.

• Create handoff context and capture learnings during checkpoints and session end.

• Keep pre-compaction hooks and session state available to Claude Code.

For non-Git projects, the same lifecycle can be invoked explicitly through the MCP tools.

Memory and search

• Store reflections and conversation context in DuckDB.

• Search by text, concept, file, project, or time-oriented queries.

• Reuse context across sessions and related repositories.

• Use local text search without an embedding service; semantic search can use a configured HTTP provider such as llama-server or Ollama and degrades gracefully when no provider is available.

Cross-project intelligence

Project groups and dependency relationships let searches include related repositories and rank results using project context. This is useful for multi-repository services, monorepos, and coordinated development work.

Quality and operational signals

Session-Buddy integrates with Crackerjack to record quality results, test patterns, failure resolutions, and workflow context. It also exposes health, Prometheus metrics, WebSocket monitoring, analytics commands, and signed skill/agent metadata for MCP clients.

Learning and skills

Session-Buddy captures reflections during checkpoints and session cleanup using deterministic extraction and content-hash deduplication. Captured knowledge can then be retrieved through the memory and search tools.

The server also publishes signed capability metadata for MCP clients:

• Skills: session_buddy_list_skills, session_buddy_get_skill

• Agents: session_buddy_list_agents, session_buddy_get_agent

These catalogs describe available capabilities; they do not perform autonomous self-modification. See Insights Capture for the capture and retrieval details.

Automatic Session Management

When the MCP server is connected from a Git repository, Session-Buddy can initialize the session on connection and perform cleanup on disconnect. The start, checkpoint, status, and end tools remain available for explicit control, and non-Git projects use that explicit workflow by default.

Lifecycle at a glance

stateDiagram-v2
    [*] --> GitRepo: Claude Code Connects
    [*] --> ManualInit: Non-Git Project

    GitRepo --> AutoStart: Auto-detect Git
    AutoStart: Initialize Session
    AutoStart --> Working: Development

    ManualInit --> ManualStart: User runs /start
    ManualStart: Initialize Session
    ManualStart --> Working: Development

    state Working {
        [*] --> Active
        Active --> Checkpoint: /checkpoint
        Checkpoint --> Active: Continue Work
        Active --> Monitoring: Track Quality
        Monitoring --> Active
    }

    Working --> AutoEnd: Disconnect/Quit
    Working --> ManualEnd: User runs /end

    AutoEnd: Auto Cleanup
    AutoEnd --> [*]: Session Handoff

    ManualEnd: Manual Cleanup
    ManualEnd --> [*]: Session Handoff

MCP Surface

The MCP server exposes a profile-gated tool surface through SESSION_BUDDY_TOOL_PROFILE:

• minimal — session lifecycle, basic search, hooks, health, baseline probes, and published agent metadata.

• standard — the daily-development surface, including conversation, extraction, knowledge graph, Crackerjack, monitoring, cross-repository, skills, and agent tools.

• full — all registered tool groups; this is the default when the variable is unset or invalid.

The active profile is defined in session_buddy/mcp/tools/profiles.py. The complete reference is in docs/user/MCP_TOOLS_REFERENCE.md.

Always-available baseline tools include:

Core session and memory tools include start, checkpoint, status, end, store_reflection, quick_search, search_summary, search_by_file, and search_by_concept.

The signed catalogs expose server-published capabilities through:

• session_buddy_list_skills and session_buddy_get_skill

• session_buddy_list_agents and session_buddy_get_agent

The HTTP service also provides /health, /healthz, and /metrics on the main service port.

Integration with Crackerjack

Crackerjack is Session-Buddy's quality and CI/CD integration point. Session- Buddy can retain quality results, test outcomes, failure patterns, and useful resolutions as session context so later checkpoints and sessions can retrieve them.

Typical local validation is:

crackerjack run --run-tests

See Crackerjack Integration for the MCP tools and integration details.

Quick Start

Prerequisites

• Python 3.14+

• uv or pip

• An MCP client that supports streamable HTTP

Install and start

git clone https://github.com/lesleslie/session-buddy.git
cd session-buddy
uv sync

# Start the streamable HTTP MCP service on 127.0.0.1:8678
uv run session-buddy server start

Useful lifecycle and diagnostics commands:

uv run session-buddy server status
uv run session-buddy server health
uv run session-buddy health
uv run session-buddy doctor

Connect an MCP client

The service endpoint is http://127.0.0.1:8678/mcp. Add an HTTP entry to the client configuration:

{
  "mcpServers": {
    "session-buddy": {
      "type": "http",
      "url": "http://127.0.0.1:8678/mcp"
    }
  }
}

Core text search works without an embedding service. Semantic search uses a configured HTTP embedding provider such as llama-server or Ollama when one is available.

Usage

After the MCP client connects, use the session prompts and tools directly:

/session-buddy:start
/session-buddy:checkpoint
/session-buddy:quick_search
/session-buddy:store_reflection
/session-buddy:end

The primary MCP tools are start, checkpoint, status, end, quick_search, search_summary, search_by_file, search_by_concept, and store_reflection. Claude Code shortcuts such as /start, /checkpoint, and /end may be generated under ~/.claude/commands/ after initialization.

Configuration

Session-Buddy uses Oneiric's layered settings model together with the repository's flat YAML compatibility layer. The project files are:

• settings/session-buddy.yaml — committed defaults

• settings/local.yaml — gitignored checkout-local overrides

• settings/lite.yaml and settings/standard.yaml — mode-specific defaults

Oneiric also checks user-level files:

• ${XDG_CONFIG_HOME:-~/.config}/session-buddy/config.yaml

• ${XDG_CONFIG_HOME:-~/.config}/session-buddy/local.yaml

Environment variables use the SESSION_BUDDY_ prefix. Nested settings use double underscores, for example:

SESSION_BUDDY_LOG_LEVEL=DEBUG
SESSION_BUDDY__DATABASE_PATH=/tmp/session-buddy.duckdb
SESSION_BUDDY_TOOL_PROFILE=standard

Runtime data defaults to ~/.claude/data/reflection.duckdb, logs to ~/.claude/logs/, and Oneiric snapshots to .oneiric_cache/ in the configured cache location.

Core session and text-search workflows do not require an external service. Embedding providers, LLM providers, and ecosystem integrations are optional and configured through the same settings and environment layers.

Memory System

Session-Buddy stores conversation context and reflections in a local DuckDB database by default. Text search, project filtering, time-aware retrieval, and reflection statistics are available locally. Semantic search is optional and uses a configured HTTP embedding provider when enabled. See Configuration for the default paths and overrides.

Session Workflow

1. Start or connect the MCP server.

2. Run /session-buddy:start when explicit initialization is needed.

3. Use /session-buddy:checkpoint during longer work sessions.

4. Search prior work with /session-buddy:quick_search or /session-buddy:search_summary.

5. Store important conclusions with /session-buddy:store_reflection.

6. Run /session-buddy:end when the session is complete.

Bodai Integration

When deployed inside the Bodai ecosystem, Session-Buddy works as the session-lifecycle and knowledge-capture layer for the Bodai components: Mahavishnu orchestration, Akosha cross-system analytics, Crackerjack quality signals, and the oneiric configuration and adapter patterns shared across components. The standalone install is unaffected — Bodai adds no special-case overrides; consumers connect through the same MCP tools and DuckDB-backed store they would in any other Claude Code environment.

Documentation

• Quick Start Guide

• Configuration Guide

• MCP Tools Reference

• Architecture Overview

• Intelligence Features

• Insights Capture

• Automatic Lifecycle

• Service Dependencies

Troubleshooting

Check the service and dependency probes first:

uv run session-buddy server status
uv run session-buddy health --json
uv run session-buddy doctor --json

If the MCP client cannot connect, confirm that the service is listening on 127.0.0.1:8678 and that the client URL ends in /mcp. Use SESSION_BUDDY_LOG_LEVEL=DEBUG for more detailed logging.

For memory or embedding issues, start with text search and then verify the configured embedding provider and its endpoint. For configuration problems, check the project YAML files, the Oneiric XDG files, and the effective SESSION_BUDDY_* environment variables.

License

BSD 3-Clause License. See LICENSE.

Acknowledgements

Session-Buddy is built on open-source foundations including FastMCP, Oneiric, mcp-common, DuckDB, Typer, and Prometheus client.