A snapshot of where humanbound/plugins is heading. This is a living document — dates and scope may change. Open a discussion or issue to weigh in.

The authoritative, continuously-updated roadmap lives at docs.humanbound.ai/plugins/.

• humanbound-test (Claude Code + Cursor) — adversarial agent testing end-to-end, with humanbound MCP integration.

• First-class LangServe detection for humanbound-test — extend the server detector to recognise langserve in pyproject.toml / requirements.txt and find add_routes(app, agent, path="/...") call sites for the entry-point hint. Highest-ROI framework miss — LangChain dominates the production AI-agent space. Files: skills/tunneling-local-agent/scripts/detect-server.py, skills/tunneling-local-agent/scripts/providers/server/langserve.sh.
• Runtime OpenAPI scrape for humanbound-test — if the detected server is already running on its configured port, hit /openapi.json and parse it directly. Framework-independent; works for anything that emits OpenAPI (FastAPI, LangServe, flask-smorest, …). Agent endpoints / payloads still live in user-authored bot-config.json — OpenAPI confirms server shape, it does not author the request body.
• Confidence scores + prompt-on-ambiguous for detection — return {fastapi: 0.95, langserve: 0.40} instead of first-match-wins, so the orchestrator can ask the user to confirm when multiple frameworks match.
• Submit humanbound-test to the official Cursor marketplace (cursor.com/marketplace/publish) so end users can install via the in-editor /add-plugin command instead of the clone+symlink flow.
• Port humanbound-tunnel (currently in the legacy humanbound-tunnel repo) — same Claude Code + Cursor dual support, add to this marketplace.

• First-class Flask + Django detection — same regex/entry-point detection approach for Flask (@app.route("/x", methods=["POST"])); Django needs manage.py show_urls (django-extensions) or urls.py parsing.
• Streamlit / Gradio adapter — these expose a UI, not HTTP endpoints. Testing them requires simulating chat via Gradio's queue API or Streamlit's session-state protocol, not just HTTP detection. Different feature, not just detection.
• TypeScript / Express / Next.js endpoint detection — needs tree-sitter or node-based parsing. Diminishing returns versus Python coverage.
• Native MCP tool gating in Claude Code via allowed-tools extended to MCP server invocations.
• Plugin templates under plugins/_template/ for community plugin authors.
• CI-based smoke tests for each plugin against a recorded MCP fixture.

• Not bundling third-party MCP servers other than Humanbound's.
• Not vendoring closed-source binaries — every plugin in this marketplace is Apache-2.0 source you can audit.
• Plugins in this marketplace may dispatch to the hosted Humanbound platform (e.g. humanbound-test uses the humanbound MCP to run adversarial tests via the backend). When a plugin requires the platform to function, it says so plainly in its README's Requirements section.

Each plugin is versioned independently via tags of the shape v-<plugin-name>-<semver>. See CONTRIBUTING.md for the release flow.