Repository avatar
Security
v1.0.0
active

sanka-mcp

io.github.sankaHQ/sanka-mcp

Sanka Desk IT requests, Sanka Flow records, approvals, workflows and app migrations, with OAuth.

Documentation

Sanka MCP Server

This repository contains the hosted Sanka MCP service and the TypeScript client it uses to call Sanka's externally callable API surface. It is now maintained as a normal TypeScript workspace, not as a Stainless-managed repository.

The production service is a remote Streamable HTTP MCP endpoint:

  • POST /mcp as the primary endpoint
  • POST /sse as a compatibility alias for clients that still expect an SSE-style path
  • POST / as a compatibility alias
  • GET /health for health checks

The /mcp endpoint exposes both the general SDK tools and the read-only CRM tools from a single MCP surface.

Live endpoints:

  • https://mcp.sanka.com/mcp
  • https://mcp.sanka.com/sse

Staging endpoint:

  • https://sanka-mcp-staging.fly.dev/mcp

Repository layout

  • src/: TypeScript client for Sanka's externally callable API surface, used by the MCP service
  • src/internal/: package-private SDK runtime helpers, not Sanka private backend code
  • packages/mcp-server/: the MCP server application and Docker entrypoint
  • .github/workflows/: CI plus staging and production Fly deployment workflows
  • fly.toml: production Fly app configuration
  • fly.staging.toml: staging Fly app configuration
  • docs/openapi-maintenance.md: guidance for keeping API coverage current without Stainless

Auth

The hosted server is a standard MCP OAuth 2.1 protected resource (MCP authorization spec 2026-07-28). Sanka (sanka-api) is the authorization server, issuer https://api-v2.sanka.com/oauth/mcp; the consent page is https://app.sanka.com/oauth/mcp/authorize.

  • Every MCP request needs Authorization: Bearer <token>. Without one the server answers 401 with WWW-Authenticate: Bearer scope="mcp:access", resource_metadata="…"; an invalid or expired token adds error="invalid_token".
  • Each endpoint is its own resource, named by the URL clients connect to: https://mcp.sanka.com/mcp, https://mcp.sanka.com/sse and https://mcp.sanka.com. Its RFC 9728 metadata is at /.well-known/oauth-protected-resource/mcp, …/sse and /.well-known/oauth-protected-resource. A token is bound to one resource.
  • /.well-known/oauth-authorization-server mirrors Sanka's metadata for clients built for the 2025-03-26 protocol.
  • Clients register with a Client ID Metadata Document (preferred) or dynamic client registration, and always use PKCE (S256); Sanka issues no client secrets.
  • The server never passes the client's token to the business API. It exchanges it through POST /oauth/internal/mcp-oauth-token (authenticated with MCP_SERVER_OAUTH_RESOURCE_SERVER_SECRET, shared with sanka-api only) for its own short-lived API token bound to the same connection. Sanka checks the connection on every API call, so disconnecting a client takes effect at once.
  • switch_workspace moves the whole connection to the chosen workspace, so later token exchanges and record URLs follow it.
  • The server keeps no MCP session: a client-sent mcp-session-id is ignored and never forwarded. Developer API keys (X-Sanka-API-Key) are refused.

The ChatGPT connector keeps its own resource at /chatgpt with its own issuer (https://api-v2.sanka.com/oauth/chatgpt).

Local development

pnpm install
pnpm build
export MCP_SERVER_OAUTH_ISSUER="http://localhost:8000/oauth/mcp"
export MCP_SERVER_INTERNAL_AUTHORIZATION_SERVER_URL="http://localhost:8000"
export MCP_SERVER_OAUTH_RESOURCE_SERVER_SECRET="<at least 32 characters>"
export MCP_SERVER_RESOURCE_URL="http://localhost:8080/mcp"
export SANKA_BASE_URL="http://localhost:8000"
node packages/mcp-server/dist/index.js --transport=http --port=8080

The local API needs the matching settings: MCP_OAUTH_ISSUER, MCP_OAUTH_AUTHORIZE_URL (the local React app's /oauth/mcp/authorize), MCP_OAUTH_RESOURCES (a JSON list with http://localhost:8080/mcp and the other local endpoints) and the same MCP_OAUTH_RESOURCE_SERVER_SECRET.

Stdio remains a local-development transport. SANKA_API_KEY there is process configuration, not a client-supplied bearer mode.

Then verify:

curl http://127.0.0.1:8080/health
curl -sS -D - http://127.0.0.1:8080/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  --data '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"curl","version":"0.1"}}}'

The initialize request answers 401 with the WWW-Authenticate challenge. Point an HTTP MCP client (for example the MCP Inspector) at http://127.0.0.1:8080/mcp without static auth headers; it discovers the authorization server, opens the Sanka consent page, and retries with its token. curl http://127.0.0.1:8080/.well-known/oauth-protected-resource/mcp shows the resource metadata.

Browser Use worker

The browser_use MCP tool dispatches allowlisted browser workflows to a separate worker. The MCP server keeps session authentication, audit metadata, workflow routing, and confirmation gates; the worker owns browser state and third-party UI interaction. The first registered workflow is:

  • demo.hubspot.company_avatar: update HubSpot demo company avatars through the HubSpot UI when CRM APIs cannot set the visible avatar.

Run a local worker with agent-browser:

npm install -g agent-browser
agent-browser install

export SANKA_BROWSER_USE_WORKER_TOKEN="local-worker-token"
export SANKA_BROWSER_USE_PROFILE_ROOT=".browser-use-profiles"
export SANKA_BROWSER_USE_ARTIFACT_DIR=".browser-use-artifacts"
pnpm --dir packages/mcp-server browser-use-worker

Point the MCP server at it:

export SANKA_BROWSER_USE_WORKER_URL="http://127.0.0.1:8787/run"
export SANKA_BROWSER_USE_WORKER_TOKEN="local-worker-token"

For production, deploy the worker as a separate service, for example on Fly, with a persistent volume mounted at SANKA_BROWSER_USE_PROFILE_ROOT so HubSpot login cookies survive restarts. Keep the worker private or token-protected and only expose /run to the MCP service.

See docs/browser-use-worker-fly.md for the Fly worker app, persistent volume, secret wiring, and profile seeding runbook.

Deployment

This repo deploys to Fly from packages/mcp-server/Dockerfile.

  • Manual production deploy: fly deploy -c fly.toml
  • Manual staging deploy: fly deploy -c fly.staging.toml
  • Automatic staging deploy: push to staging
  • Automatic production deploy: publish a GitHub release that targets main

Release Flow

Use the same promotion shape as the main Sanka app:

  1. Open feature PRs into staging.
  2. Merge staging after CI passes to deploy sanka-mcp-staging.
  3. Validate the staging MCP endpoint.
  4. Open staging -> main.
  5. After main is ready, run Create new Sanka MCP Tag and Release.
  6. The published release deploys production.

Required Fly apps and secrets:

  • Production app: sanka-mcp
  • Staging app: sanka-mcp-staging
  • Set the same MCP_SERVER_* secrets on both apps unless staging intentionally uses a different auth stack.

Retry policy

The TypeScript client retries 408, 409, 429, 5xx and connection errors twice by default. Only tools with metadata.operation: 'read' keep those automatic retries. Every write tool calls the API once and returns any error to the agent: a write can fail after its side effect happened, such as an email Gmail accepted before Sanka recorded it, and a retry would repeat it. The API also marks errors from delivery routes with x-should-retry: false. Reply tools report DELIVERY_UNKNOWN as a do-not-resend error.

Maintenance direction

This repository no longer depends on Stainless project access at runtime or for ongoing development. API coverage should be maintained directly in this repo.

The recommended next step is to adopt open-source OpenAPI tooling for updates to the TypeScript client, rather than reintroducing a hosted generator dependency. See openapi-maintenance.md.

The repo now includes a starter typegen command:

pnpm generate:openapi-types

By default it reads the sibling Sanka spec at ../sanka-sdks/openapi.json.

Developer Cloud (release candidate)

The hosted tools expose the same V2 source, run, receipt, certificate and Fleet contracts as the CLI and SDKs. This source change does not enable production execution. Read get_developer_cloud_availability first and obey each capability's availability flag. Free local migration and client-evidence registration are unchanged.

Every call takes a pinned workspace_id. Paid creates and selective retries require an explicit request, approved cap, confirm: true, and stable idempotency_key. Existing user authorization is sufficient. A lost response must not create a new intent: read back or replay the identical key and request. No tool purchases credits, creates repository PRs, merges, or deploys.

  • Source/run: upload_developer_cloud_source, create_developer_cloud_run, list_developer_cloud_runs, get_developer_cloud_run, cancel_developer_cloud_run.
  • Evidence: list_developer_cloud_events, get_developer_cloud_receipt, list_developer_cloud_artifacts, download_developer_cloud_artifact.
  • Certificates: list_developer_cloud_certificate_keys, get_developer_cloud_certificate, revoke_developer_cloud_certificate.
  • Fleet: create_developer_cloud_fleet, list_developer_cloud_fleets, get_developer_cloud_fleet, cancel_developer_cloud_fleet, retry_developer_cloud_fleet.

Repair and certificate issuance use create_developer_cloud_run.request.repair and .certification, respectively. Compute is 25 credits per worker-minute, rounded once per run. Successful Repair adds 250 credits; independent certificate issuance adds 500 (1 credit = ¥1 or 1 cent since 2026-11-20). All compute and premiums remain within the approved cap. Failed gates have no premium. Fleet adds no surcharge: its cap is the sum of explicit child caps, reserved atomically, with at most five active workers per workspace. Retry selects only failed children after the original Fleet settles; it creates a new Fleet and preserves completed children's receipts.

Upload the exact ZIP digest and declared full revision; the service pins uploaded bytes but does not independently prove their Git provenance. Inspect certificate scope, scenarios, limitations and revocation. Downloading a certificate does not verify its signature: use the CLI's offline verification with the published key. Artifact downloads verify bytes against recorded SHA-256 metadata. MCP supports up to 64 MiB; larger artifacts use the CLI/API. Large downloads use the existing session-bound URL or read_binary_download_chunk flow.

Request-schema regeneration uses the shared V2 SDK input:

python3 scripts/sync-developer-cloud-contract.py ../sanka-sdks/openapi.json
pnpm exec prettier --write packages/mcp-server/src/generated/developer-cloud-schemas.ts

Embedded workspace

The Sanka app opens the connected full Flow workspace: its sidebar, dashboards, record tables, workflows and reports. Full workspace access requires explicit Sanka authorization. Connections with narrower access show the authorization controls; the earlier Invoice assistant pilot screen is no longer offered.

The hosted invoice MCP tools remain available under their existing scopes and confirmation requirements. Removing the pilot screen does not change invoice records or the API contract.

No installation packages available.

Remote