MCP Context Provider

Persistent context and learned instincts for Claude Desktop and Claude Code — surviving across sessions.

A TypeScript MCP server that gives Claude persistent Contexts (static tool rules) and Instincts (learned, confidence-scored rules distilled from sessions). No more re-establishing context in every new chat.

Two core concepts:

Four subsystems:

• Engine — loads, matches, and merges contexts + instincts into injection payloads
• MCP Server (src/server/index.ts) — stdio + HTTP transport, 6 MCP tools
• CLI (mcp-cp) — approval registry for instinct lifecycle management
• Memory Bridge — optional sync of instincts to mcp-memory-service

git clone https://github.com/doobidoo/MCP-Context-Provider.git
cd MCP-Context-Provider
npm install
npm run build

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "context-provider": {
      "command": "node",
      "args": ["/path/to/mcp-context-provider/dist/server/index.js"],
      "env": {
        "CONTEXTS_PATH": "/path/to/mcp-context-provider/contexts",
        "INSTINCTS_PATH": "/path/to/mcp-context-provider/instincts"
      }
    }
  }
}

Add to ~/.mcp.json:

{
  "mcpServers": {
    "context-provider": {
      "command": "node",
      "args": ["/path/to/mcp-context-provider/dist/server/index.js"],
      "env": {
        "CONTEXTS_PATH": "/path/to/mcp-context-provider/contexts",
        "INSTINCTS_PATH": "/path/to/mcp-context-provider/instincts"
      }
    }
  }
}

Important: Use absolute paths for both args and env values. Claude Code does not support the cwd field in MCP server configs — relative paths will resolve from the wrong directory and the server will fail to connect.

Install directly from the marketplace:

/plugin marketplace add doobidoo/MCP-Context-Provider
/plugin install context-provider

This auto-configures the MCP server with correct paths — no manual .mcp.json editing needed.

Install the skill globally (stays current with git pull):

mkdir -p ~/.claude/skills/instill
ln -s /path/to/mcp-context-provider/.claude/skills/instill.md ~/.claude/skills/instill/SKILL.md

Then use /instill at the end of productive sessions to distill learned patterns into instinct candidates.

The instill-trigger hook automatically detects mistakes during a session and nudges Claude to suggest /instill when a threshold is reached. It monitors:

• User corrections (UserPromptSubmit) — "no not that", "that's wrong", "still broken", etc.
• Tool failures (PostToolUse) — non-zero exit codes, tracebacks, permission errors

Install the hook:

cp hooks/instill-trigger.js ~/.claude/hooks/core/instill-trigger.js

Register in ~/.claude/settings.json under both UserPromptSubmit and PostToolUse:

{
  "type": "command",
  "command": "node --no-warnings \"~/.claude/hooks/core/instill-trigger.js\"",
  "timeout": 3
}

Scoring: Corrections weighted 1.5x, tool failures 0.5x. Combined threshold: 3.0. Max 1 nudge per session. All tunable via CONFIG object in the hook file.

Contexts are JSON files in contexts/*_context.json. Each file matches one or more tools via glob patterns and injects static rules.

{
  "tool_category": "git",
  "description": "Git workflow rules",
  "auto_convert": false,
  "metadata": {
    "version": "1.0.0",
    "applies_to_tools": ["git:*", "Bash"],
    "priority": "high"
  },
  "syntax_rules": { ... },
  "auto_corrections": {
    "fix-1": { "pattern": "...", "replacement": "..." }
  }
}

Add a new context by dropping a *_context.json file in contexts/ and restarting the server.

Instincts are YAML files in instincts/*.instincts.yaml. They are distilled from sessions via /instill and require human approval.

version: "1.0"

instincts:
  my-rule:
    id: my-rule
    rule: "Compact, actionable rule (20–80 tokens)."
    domain: git
    tags: [git, workflow]
    trigger_patterns:
      - "git commit"
    confidence: 0.75
    min_confidence: 0.5
    approved_by: human
    active: true
    created_at: "2026-03-10T00:00:00Z"
    outcome_log: []

Manage instincts with the CLI:

mcp-cp list
mcp-cp show <id>
mcp-cp approve <id>
mcp-cp reject <id>
mcp-cp tune <id> --confidence 0.8
mcp-cp outcome <id> + "worked well"

npm run build     # Compile TypeScript
npm run dev       # Watch mode
npm run lint      # Type-check only
npm test          # Run tests (vitest)
npm start         # stdio transport
npm run start:http  # HTTP transport on port 3100

No. /instill is a Claude Code skill (.claude/skills/instill.md) and only works in the Claude Code CLI. Claude Desktop does not have a skill system.

However, you can achieve the same result in Claude Desktop:

1. MCP tools work in both - The list_instincts and build_injection tools are available in Claude Desktop via the MCP server.
2. For the instill workflow, create a Claude Desktop Project and paste the instill instructions as Custom Instructions. Claude Desktop can then use desktop-commander or similar MCP servers to write YAML files.

The reason /instill is not exposed as an MCP tool: it is an interactive, multi-step workflow (analyze conversation, present candidates, await user decision, write YAML). MCP tools return a single response and cannot drive multi-turn interactions.

Potentially yes. Instincts distilled from work sessions may contain internal hostnames, customer names, infrastructure details, or operational procedures. The learned.instincts.yaml file is tracked by git by default, so review its contents before pushing to public repositories. Consider adding it to .gitignore if your instincts contain proprietary information.

See CHANGELOG.md.

MIT