codeany-hubs-mcp-server 0.0.3.3

MCP server exposing Codeany Hubs authoring tools.

CodeAny Hubs MCP Server

This repository packages a Model Context Protocol (MCP) JSON-RPC server that exposes the full CodeAny Hub task authoring surface area. It builds on the official codeany-hub Python SDK without modifying it, wiring in consent policies, token storage rules, and per-tool routing so that IDE assistants can author and manage tasks securely.

Features

• Session and destructive-action consent prompts with configurable messaging.
• Hub allowlist filtering and token store controls (memory or file-backed).
• Task CRUD, limits, statements, IO/checker, testsets, per-test management, and examples tools.
• Multiple Choice Question (MCQ) tooling to create/convert tasks and manage question configs directly from MCP clients.
• Streaming support for large uploads (e.g., ZIP testsets) with size enforcement.
• Drop-in CLI entrypoint codeany-hubs-mcp for stdio transports.
• MCP-spec compatibility: spec-compliant initialize (includes protocolVersion) and standard endpoints tools/list & tools/call (plus prompts/list, resources/list aliases).

Quickstart

1. Install the package

   pip install codeany-hubs-mcp-server
   

2. Set required environment

   export CODEANY_API_TOKEN="your-codeany-hub-token"
   export MCP_ASK_ON_START=true
   export MCP_TOKEN_STORE_MODE=memory  # or 'file'
   

   The SDK also accepts other credential styles—CODEANY_USERNAME/CODEANY_PASSWORD, device codes, or JWTs. Use whichever combination you already rely on; the MCP server reuses the SDK's authentication logic.

3. Run the MCP server

   codeany-hubs-mcp
   

   The server listens on stdio. Pair it with your MCP-compatible client or IDE integration.

Documentation for configuration flags, security posture, and tool schemas lives under docs/.

Configuration

MCPServerConfig is loaded from environment variables prefixed with MCP_. Highlights:

Variable	Description	Default
MCP_ASK_ON_START	Prompt once per session for consent	true
MCP_ASK_ON_DESTRUCTIVE_OPS	Reconfirm before mutating actions	true
MCP_TOKEN_STORE_MODE	memory or file token persistence	memory
MCP_TOKEN_STORE_PATH	File path for tokens (required if mode=file)	None
MCP_HUB_ALLOWLIST	Comma separated hub slugs allowed	None (all)
MCP_MAX_UPLOAD_MB	Max upload size across tools	25
MCP_MAX_EXAMPLES	Cap on examples saved per task	50
MCP_ALLOW_LOCAL_PATHS	Permit reading local files for uploads	false
MCP_NONINTERACTIVE_CONSENT	Auto-answer session prompts when stdin isn't a TTY	true
MCP_NONINTERACTIVE_CONSENT_DESTRUCTIVE	Auto-answer destructive prompts when stdin isn't a TTY	false
MCP_LOG_PATH	File path for server logs (None to disable file logging)	"codeany-hub-mcp.log"
MCP_LOG_LEVEL	Logging verbosity (CRITICAL, ERROR, WARNING, INFO, DEBUG)	INFO

Refer to docs/getting-started.md for the full list and examples.

Tool Catalog

All handlers register under the following namespaces:

• hubs.* for discovery.
• tasks.* for task lifecycle, limits, statements, IO/checker, testsets, tests, and examples.
• tasks.mcq.* for Multiple Choice Question configuration (question text, options, correctness flags).
• tools.capabilities returns the above registry.
• Tools are exposed under both a sanitized name (for example tasks_list) and their original dotted name (for example tasks.list); tools/call accepts either form, but not UI decorations such as tasks_create (codeany-hubs MCP Server).
• Many tasks.* tools, including tasks.create, are thin wrappers around the CodeAny Hub SDK. The MCP server forwards all arguments except confirm directly into SDK methods such as client.tasks.create. For these tools, consult the SDK docs for the full parameter list and avoid inventing new fields.

Each handler returns plain JSON dictionaries (obtained via model_dump() on SDK models). See docs/tools-reference.md for parameter and result schemas.

Development

Run linting and tests via:

scripts/run_tests.sh

The script runs Ruff, MyPy, and pytest. Tests rely on lightweight stubs instead of live HTTP calls to keep the suite fast and deterministic.

Development Checklist

• Install dev dependencies: pip install -e '.[dev]' (or update your active venv).
• Export/verify required env vars (CODEANY_*, MCP_*) or load from .env.
• Run scripts/run_tests.sh before pushing (Ruff → MyPy → pytest).
• Update docs when adding or changing tool signatures (docs/tools-reference.md).
• Validate security posture: consent prompts flow, MCP_TOKEN_STORE_MODE defaults to memory, no secrets committed, and logging redacts sensitive fields.
• Regenerate roadmap/status updates in README.md and AGENTS.md when milestones shift.
• For releases: bump version in pyproject.toml, tag, and publish to PyPI/registry.

Roadmap & TODOs

• Transport Options
  • Deliver a WebSocket gateway so IDEs can connect over persistent channels (target Q4 2024).
  • Ship a reverse proxy recipe for MCP-over-HTTP to simplify remote deployments.
• Schema & Tooling
  • Publish JSON Schema definitions for every tool payload (param/result) to unblock code generation.
  • Add auto-generated OpenAPI-style docs mirroring the schema to aid documentation sites.
• Testing & Quality
  • Stand up a mocked CodeAny Hub API harness for deterministic integration tests.
  • Add contract tests ensuring server responses align with SDK model versions.
  • Automate regression suites in CI (GitHub Actions) with matrix runs across Python 3.9-3.12.
• Async & Performance
  • Expose async handler variants when the SDK’s async clients exit beta.
  • Implement configurable concurrency limits and request queue metrics.
• Deployment & Ops
  • Publish container images (ghcr.io/codeany/codeany-hubs-mcp-server) with baked-in health endpoints.
  • Provide Terraform/Helm snippets for managed rollouts.
  • Integrate structured logging (JSON) and optional OpenTelemetry exporters for tracing.

Security Highlights

• Consent prompts are enforced centrally (MCPAuthorizationGuard).
• Transport-safe prompts: the server never reads from STDIN (which carries JSON-RPC). It uses the controlling TTY when available. In headless environments, session consent is auto-accepted and destructive prompts are auto-rejected (clients must pass confirm=true), preventing transport corruption and working reliably across IDEs.
• Token storage defaults to in-memory; file-backed storage must be explicitly configured.
• Upload helpers coerce data URIs, byte blobs, or (optionally) vetted local paths while enforcing size limits.
• No request bodies or secrets are logged unless MCP_LOG_REQUESTS=true, which only prints method/path/status.

Logging & Diagnostics

• On startup, the server writes a readiness message to stderr indicating the active log destination.
• File logs default to codeany-hub-mcp.log; adjust via MCP_LOG_PATH or disable by setting it to an empty value.
• Set MCP_LOG_LEVEL to DEBUG when troubleshooting; leave at INFO in production.
• Router-level debug logs capture incoming requests and streaming progress while avoiding payload contents.

IDE / Tool Integration

The MCP server communicates via stdio by default. Follow the platform-specific steps below to wire it into your tooling:

• Codex CLI / VS Code (Codex extension)

  1. Ensure the package is installed in your project environment.
  2. Add the following entry to your Codex configuration (.codex/config.json or settings UI):

     {
       "name": "codeany-hubs",
       "command": "codeany-hubs-mcp",
       "cwd": "/path/to/your/project",
       "env": {
         "CODEANY_API_TOKEN": "…",
         "MCP_TOKEN_STORE_MODE": "memory"
       }
     }
     

  3. Restart Codex; the tool list should now include codeany-hubs under MCP providers.

• Claude Code / Cursor

  1. Open the MCP server panel (Claude → Settings → MCP Servers).
  2. Add a new server with command codeany-hubs-mcp and inherit your workspace environment.
  3. For remote runs, point to a wrapper script that exports the required env vars before launching the server.
  4. Save and reconnect; Claude Code will request consent via the prompt hook when first used.

• Other MCP-compatible tools

  • Provide the executable (codeany-hubs-mcp) as the server command.
  • Ensure stdin/stdout remain unbuffered (most frameworks handle this automatically).
  • For GUI clients, implement the consent prompt callback to display the text passed from the server.

Tip: keep a dedicated .env file with both CODEANY_* and MCP_* variables and load it before spawning the server to maintain consistent environments across tools.

Non-interactive hosts (Cursor / Claude Code / JetBrains / others): The server avoids reading prompts from stdin when it isn’t a TTY to preserve JSON‑RPC framing. Control answers via:

• MCP_NONINTERACTIVE_CONSENT (defaults to true) for session prompts.
• MCP_NONINTERACTIVE_CONSENT_DESTRUCTIVE (defaults to false) for destructive prompts.

For scripted/automated flows, prefer passing confirm=true on destructive tool calls to skip prompts entirely.

MCP Usage Instructions for AI Agents

AI assistants that consume this MCP server should follow the same operating rules captured in AGENTS.md. In short:

1. Always load the environment contract

   • Read AGENTS.md before issuing tool calls; it documents required env vars, consent expectations, and the full dev checklist.
   • Verify CODEANY_* credentials plus MCP_* toggles (token store mode, allowlist, upload limits) are set in the session.

2. Respect consent + confirmation semantics

   • Invoke guard.ensure_session_consent once per session (the server does this automatically, but agents should surface prompts to users).
   • Pass confirm=true for destructive tools when the operator already approved the action (delete, rename, update, etc.).

3. Use documented tool schemas

   • Parameters and responses are defined in docs/tools-reference.md. Only send fields listed there; pass task references as slugs or IDs consistently.
   • For thin wrappers around the SDK (for example tasks.create), this MCP server does not reshape arguments: it calls the underlying SDK client (for tasks.create this is client.tasks.create) with your parameters, except that confirm is stripped before the call. Always use exactly the field names supported by the SDK and do not add extra top-level keys like name or task_type unless the SDK documents them.
   • Prefer tasks.statements.update for statement mutations (instead of legacy upsert calls); include statement_id and the desired payload.

4. Handle errors deterministically

   • All failures surface structured errors.to_mcp_error payloads. Detect consent_rejected, hub_not_allowed, validation_error, etc., and relay actionable guidance to the user.

5. Keep README and AGENTS.md in sync

   • If you edit the roadmap, checklist, or security posture for agent consumption, update both files so every AI client shares the same playbook.

Following these steps ensures any AI agent—Codex, Claude Code, Cursor, or future MCP clients—uses the server safely and consistently. Refer to AGENTS.md for the canonical format of responsibilities, checklists, and escalation procedures.

Additional Resources

• docs/getting-started.md
• docs/security.md
• docs/tools-reference.md
• docs/streaming-and-uploads.md