> For the complete documentation index, see [llms.txt](https://docs.openalgo.in/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.openalgo.in/change-log/release/version-2.0.1.0-released.md). # Version 2.0.1.0 Released **Date: 3rd May 2026** **Major Feature Release: Remote MCP — Self-Hosted OAuth 2.1 + MCP HTTP/SSE Server for ChatGPT, Claude.ai, and Claude Mobile, Plus Per-Purpose 2FA Enforcement, Symbol Search Expansion, Admin Diagnostics, and Zerodha NCO/GLOBAL\_INDEX Support** This is the biggest release in the 2.0.x line, covering **20+ commits** since v2.0.0.9. The headline change is **Remote MCP** — a self-hosted OAuth 2.1 + Model Context Protocol HTTP/SSE transport that lets hosted AI clients (ChatGPT.com, Claude.ai, Claude iOS / Android) connect to your OpenAlgo install over HTTPS with the same 40 tools the local stdio MCP already exposes. The local stdio MCP (Claude Desktop, Cursor, Windsurf) is **untouched** — Remote MCP is a parallel, opt-in transport, off by default, that ships behind two enabler scripts (one for native Ubuntu, one for Docker) and a full admin operations console at `/admin/remote-mcp`. Alongside Remote MCP, this release lands per-purpose TOTP enforcement (login / MCP / password reset), a multi-exchange Symbol Search expansion (issue #1326), an admin Diagnostics page with downloadable system reports, and Zerodha NCO + GLOBAL\_INDEX exchange support. *** **Highlights** * **Remote MCP — Hosted AI Clients via OAuth** — Brand-new `/mcp` HTTP/SSE transport with full OAuth 2.1 + PKCE, Dynamic Client Registration (DCR), JWKS-published RS256 JWTs, refresh-token rotation with reuse-detection family revocation, and per-token-per-scope rate limits. ChatGPT-compatible discovery (`/.well-known/oauth-authorization-server`, `/.well-known/oauth-protected-resource`), claude.ai-compatible scope flow. Off by default; turned on by `install/enable-remote-mcp.sh` (native) or `install/enable-remote-mcp-docker.sh` (Docker). * **Per-purpose 2FA enforcement** — TOTP can now be required independently for **dashboard login**, **MCP authorization (`write:orders` consent)**, and **password reset** — set the master switch in Profile → TOTP, then pick which purposes apply. Saving requires a fresh TOTP code in the same request to prove authenticator access for both enabling and disabling. * **Admin → Remote MCP operations console** — Approve / revoke DCR clients, browse the full MCP tool-call audit log (`log/mcp.jsonl`) by tool / scope / outcome, and a one-click **Kill switch** that atomically revokes every refresh token across every approved client. * **Symbol Search rewrite (#1326)** — Multi-exchange + multi-instrumenttype filtering, CSV download, AmiBroker / TradingView / Python / Excel copy formats, lifted the 500-row hard cap, exchange-only browse mode, and per-user search history. * **Admin Diagnostics page** — `/admin/diagnostics` shows live system info (Python / Flask versions, DB sizes, broker session state), a paginated error browser over `log/errors.jsonl`, and a one-click downloadable diagnostic bundle (env-redacted) for support tickets. * **Zerodha NCO + GLOBAL\_INDEX exchanges** — NCO (NSE Commodities) and GLOBAL\_INDEX (US30 / JAPAN225 / HANGSENG, plus GIFTNIFTY from NSE IFSC) now route correctly through the master contract download and quote endpoints. Quote-only on GLOBAL\_INDEX (no orders by exchange convention). * **"Virtual / Paper" → "Sandbox" rename** — All in-product copy, docs, and SDK examples now uniformly say *Sandbox* (the database, blueprint, and API endpoint names were already `sandbox` — only display strings were inconsistent). * **Platform version bump** — `2.0.0.9` → `2.0.1.0`. `.sample.env` `ENV_CONFIG_VERSION` `1.0.6` → `1.0.7`. SDK pin (`openalgo==1.0.49`) unchanged. *** **Remote MCP — feature deep dive** Remote MCP brings the OpenAlgo MCP toolset to **hosted AI clients** over the public internet. Same 40 tools as the local stdio integration, exposed at `https://yourdomain.com/mcp` with full OAuth 2.1. **Architecture in one sentence:** the hosted client (ChatGPT / Claude.ai) only ever holds an OAuth Bearer JWT signed by your server's RS256 keypair; tool dispatch on the server side reuses your existing `/api/v1/*` API key (looked up server-side at boot) over loopback. The hosted client never sees your API key or your broker tokens. **OAuth foundation (Phase 2a + 2c + 2d)** — `e86cf450f`, `0f4f71f0`-derived family, `cb9350f27`: * RS256 keypair generated and rotated via `utils/oauth_keys.py`; public set published at `/oauth/jwks.json` * Dynamic Client Registration at `/oauth/register` (RFC 7591) — clients land in a *pending* bucket, gated behind admin approval (default `MCP_OAUTH_REQUIRE_APPROVAL=True`) * Authorization Code + PKCE-S256 only — `plain` is not advertised; `alg=none` and `alg=HS256` JWTs are rejected by the verifier * Refresh-token rotation with **family revocation** (RFC 6749 §10.4): single-use, atomic `UPDATE ... WHERE revoked_at IS NULL`, replay of a revoked token revokes the entire family * `/oauth/revoke` for explicit token retirement, `/oauth/token` for code-exchange and refresh * Argon2 hashing with `API_KEY_PEPPER` for client\_secret + refresh-token storage in `database/oauth_db.py` **Per-purpose 2FA (Phase 2b + Phase 2 UI)** — `dbc595e88`, `ff44adf46`: * Four new boolean columns on `users`: `totp_enabled` + `totp_required_for_login` + `totp_required_for_mcp` + `totp_required_for_password_reset` * `is_totp_required_for(purpose)` helper centralizes the gating decision * Login flow: POST `/auth/login` → on TOTP requirement, returns `totp_required` flag + temp ticket → POST `/auth/login/totp` with code * Configure flow at `/auth/2fa/configure` requires a fresh TOTP in the same request (proves authenticator access for *both* enabling and disabling — closes the "stolen session can disable 2FA" hole) * React UI: `TwoFactorEnforcement.tsx` profile-page toggle, `Login.tsx` TOTP step, `ResetPassword.tsx` TOTP path **HTTP transport (Phase 3)** — `7d805af15`: * JSON-RPC 2.0 dispatcher at `/mcp` (POST) + Server-Sent Events transport for streaming * Per-token-per-scope sliding-window rate limits (`MCP_RATE_LIMIT_READ` / `MCP_RATE_LIMIT_WRITE`) * CORS allowlist (default: `https://claude.ai,https://chatgpt.com`) with `WWW-Authenticate` exposed via `Access-Control-Expose-Headers` so browsers can read the realm hint on 401 * Audit log writes to `log/mcp.jsonl` — every tool call with timestamp, JTI, scope, outcome, latency, `params_hash` (raw params are deliberately *not* logged) * Tool registry (`utils/mcp_tool_registry.py`) maps all 40 tools to scopes (`read:market` / `read:account` / `write:orders`) and exposes FastMCP-generated JSON schemas in the `tools/list` response * Pre-flight refusal: server refuses to boot with `MCP_HTTP_ENABLED=True` and `FLASK_DEBUG=True` together (debug tracebacks would leak bearer tokens) **Install integration (Phase 4)** — `b36637b5e`, `90897a550`: * `install/enable-remote-mcp.sh` — native Ubuntu enabler. Detects all `openalgo-*` systemd services, refuses if `FLASK_DEBUG=True`, backs up `.env`, sets the four MCP keys, runs `upgrade/migrate_all.py`, restarts the service, and probes the discovery / JWKS / `/mcp/healthz` endpoints to confirm boot. * `install/enable-remote-mcp-docker.sh` — Docker enabler. Walks `/opt/openalgo/*/docker-compose.yaml`, picks a stack, backs up the bind-mounted `.env`, updates keys, restarts the container (whose `start.sh` runs migrations automatically), and runs the same smoke probe. * `install/Remote-MCP-readme.md` — operator-focused install guide with same-domain Mode 1 (automated) and subdomain Mode 2 (manual nginx + certbot recipe), threat model, and disabling instructions. * New `docs/userguide/remote-mcp.md` — end-user guide for connecting ChatGPT (Settings → Apps → New App BETA → Advanced OAuth → DCR) and Claude.ai (Settings → Connectors → + → Add custom connector). **Admin operations (Phase 5)** — `8be942a7c`: * `/admin/remote-mcp` React page with three tables (Pending / Approved / Revoked) + audit viewer + kill switch * `GET /admin/api/oauth/clients` lists clients by status; approve/revoke endpoints require typed-string confirmation for destructive actions * `GET /admin/api/mcp/audit` — paginated audit log over `log/mcp.jsonl` with whitelisted query keys * `POST /admin/api/mcp/kill-switch` — typed-string confirm, atomically revokes every refresh token across every approved client (read-only access tokens still expire on their existing 15-minute TTL) **Security audit fixes + ChatGPT compatibility** — `926a597bf`: * Migrated from deprecated `authlib.jose` → `joserfc` with explicit `algorithms=["RS256"]` pinning * `MCP_PUBLIC_URL` is a hard requirement when `MCP_HTTP_ENABLED=True` — collapsing it to empty would let JWTs minted on instance A be replayed against instance B's loopback * `error_detail` in MCP responses replaced with generic *"Tool execution failed"* — the full detail is in the audit log only, so SQL errors / stack traces don't leak to the model * Added path-relative discovery alias `/mcp/.well-known/oauth-protected-resource` because ChatGPT fetches it that way (RFC 9728 says root-relative is canonical, but ChatGPT does both) * Default `MCP_RATE_LIMIT_WRITE` raised from 5/min → 50/min based on real ChatGPT/Claude usage patterns * Removed the vestigial `MCP_OAUTH_LOGIN_AUTH_LEVEL` env var (replaced by the per-purpose 2FA flags) **CSP and consent screen fixes** — `52c96a11a`, `5f60b4817`, `3175a4104`: * Per-page CSP on the consent screen sets `form-action` to allow exactly the registered redirect\_uri's origin — fixes the form-submit block when the global CSP middleware would otherwise refuse the cross-origin POST * Global CSP middleware (`csp.py`) now respects view-set CSP headers (won't overwrite if a header is already set) * `tools/list` response now includes the real Pydantic-generated JSON schemas (reach into FastMCP's `_tool_manager._tools`) so ChatGPT stops hallucinating parameter names like `product_type` instead of `product` * Removed Jinja-time `csrf_token()` dependency on the consent template — render the token via `_csrf_token_value()` helper inside the view so timing-of-globals isn't an issue **Three new admin endpoints + four new database models:** | New | Purpose | | ----------------------- | -------------------------------------------------------------------------------------- | | `database/oauth_db.py` | `OAuthClient`, `OAuthRefreshToken` (with `family_id` / `parent_id`), `OAuthSigningKey` | | `database/user_db.py` | 4 boolean columns + `find_user_by_exact_username()` | | `utils/oauth_keys.py` | RS256 keypair generation + rotation + `public_jwks()` | | `utils/oauth_tokens.py` | `issue_access_token`, `rotate_refresh_token`, `verify_access_token` | **Default security posture:** | Setting | Default | Why | | ------------------------------- | ------------ | ----------------------------------------------------- | | `MCP_HTTP_ENABLED` | `False` | Off until you opt in | | `MCP_OAUTH_REQUIRE_APPROVAL` | `True` | DCR clients land pending until admin approves | | `MCP_OAUTH_WRITE_SCOPE_ENABLED` | `False` | Order placement unreachable until you flip the switch | | `MCP_RATE_LIMIT_READ` | `60/min` | Per-token sliding window | | `MCP_RATE_LIMIT_WRITE` | `50/min` | Per-token sliding window | | `MCP_MAX_ORDER_QTY` | `0` (no cap) | Recommend setting a sane cap | *** **Symbol Search expansion (#1326)** * `232c637fb` — `feat(search): lift 500 cap, allow exchange-only browse, add search history` * `f0c03eede` — `feat(search): multi-exchange/instrumenttype, CSV download, copy formats` The Symbol Search page now supports filtering across multiple exchanges and multiple instrument types in a single query, browsing an entire exchange without entering a search term, downloading the result set as CSV, and copying selections in AmiBroker / TradingView / Python / Excel formats. The previous 500-row hard cap is gone — large result sets stream incrementally. Per-user search history is preserved across sessions. *** **Admin / Operations** * `566113d49` — `feat(admin): add Diagnostics page with system info, error browser, and downloadable report` `/admin/diagnostics` consolidates the moving parts of "what's going on with this install" into a single React page: Python / Flask / SQLAlchemy versions, database sizes for all six SQLite/DuckDB databases, broker session state, configured env vars (with secrets redacted), a paginated browser over `log/errors.jsonl` with stack traces and Flask request context, and a one-click **Download diagnostic bundle** that produces an env-redacted ZIP suitable for attaching to a support ticket. *** **Brokers** **Zerodha** * `6bc37381e` — `feat(zerodha): support NCO and GLOBAL_INDEX exchanges` Adds two Zerodha-only exchange codes: * **NCO** — NSE Commodities (Zerodha's symbol format differs from NCDEX; this fix routes correctly through the master contract download) * **GLOBAL\_INDEX** — US30, JAPAN225, HANGSENG, plus `GIFTNIFTY` from NSE IFSC. Quote-only by exchange convention; orders are not supported on GLOBAL\_INDEX. *** **Documentation** * `199c544d1` — `docs: rename "virtual/paper trading" to "sandbox trading" terminology` * `6efaf1655` — `docs: rename remaining "virtual" trading terms to "sandbox" equivalents` * `500e27cbb` — `docs: add Remote MCP user guide for ChatGPT and Claude.ai` * New `install/Remote-MCP-readme.md` — operator-focused install + threat model * New `docs/prd/remote-mcp.md` — full product requirements doc with architecture, MUST/SHOULD/COULD security controls The "virtual / paper trading" rename only touched display strings — the database (`db/sandbox.db`), blueprint (`blueprints/sandbox.py`), and API endpoints (`/api/v1/sandbox/*`) were already named `sandbox`. The rename closes a long-standing inconsistency between in-product copy and the underlying schema. *** **Configuration changes** `.sample.env`: * `ENV_CONFIG_VERSION` `1.0.6` → `1.0.7` * New section: `MCP_HTTP_ENABLED`, `MCP_PUBLIC_URL`, `MCP_OAUTH_REQUIRE_APPROVAL`, `MCP_OAUTH_WRITE_SCOPE_ENABLED`, `MCP_HTTP_CORS_ORIGINS`, `MCP_HTTP_IP_ALLOWLIST`, `MCP_OAUTH_ACCESS_TTL`, `MCP_OAUTH_REFRESH_TTL`, `MCP_OAUTH_CODE_TTL`, `MCP_RATE_LIMIT_READ`, `MCP_RATE_LIMIT_WRITE`, `MCP_MAX_ORDER_QTY` * Removed: `MCP_OAUTH_LOGIN_AUTH_LEVEL` (vestigial — replaced by per-purpose 2FA flags) `pyproject.toml`: * `version = "2.0.1.0"` * SDK pin (`openalgo==1.0.49`) unchanged `utils/version.py`: * `VERSION = "2.0.1.0"` *** **Upgrade procedure** **For existing installs (Native Ubuntu):** ```bash cd /var/python/openalgo-flask//openalgo sudo ./install/update.sh # update.sh runs migrate_all.py — schema changes for the OAuth + 2FA columns # are applied automatically. Remote MCP stays disabled by default. ``` **For existing installs (Docker):** ```bash cd /opt/openalgo/ sudo docker compose pull sudo docker compose up -d # The container's start.sh runs migrate_all.py before gunicorn boots. # Remote MCP stays disabled by default. ``` **To enable Remote MCP (after upgrading):** ```bash # Native Ubuntu sudo ./install/enable-remote-mcp.sh # Docker sudo ./install/enable-remote-mcp-docker.sh ``` Both enabler scripts are idempotent and back up `.env` before any change. They print a one-liner rollback command if the smoke probe fails. **For local developers (uv):** ```bash git pull origin main uv sync cd frontend && npm install && npm run build uv run app.py ``` *** **Contributors** * **@marketcalls (Rajandran)** — release management, Remote MCP architecture and full OAuth 2.1 implementation (5 phases — scaffold → OAuth foundation → 2FA wiring → discovery + JWKS + DCR → /authorize + /token + /revoke → Phase 2 UI → HTTP/SSE transport → install integration → admin UI), security audit + joserfc migration, CSP fixes for the consent flow, ChatGPT-compatibility hardening (path-relative discovery, real tool schemas, generic error\_detail), Docker enabler, install integration, comprehensive PRD + operator docs + end-user docs, Symbol Search rewrite (#1326), Admin Diagnostics page, Zerodha NCO + GLOBAL\_INDEX support, and the "virtual → sandbox" terminology cleanup. *** **Links** * **Repository**: * **Documentation**: * **Remote MCP guide**: * **Discord**: * **YouTube**: * **Issue tracker**: ***