Aller au contenu principal

Feature F — MCP Connections (manual servers + bindings + tool funnel)

Branch: session/feat-mcp-connections. Implements the MCP slice of the merged docs/specs/features/agent-plugins/ spec (plan §2.4/§2.5, tasks T23–T27 adapted) plus the founder-added manual connection source that works without the package system (T1–T22 remain out of scope).

What shipped​

  • Manual MCP connection registry (mcp_server_connections): user-global rows with name (slug-safe, becomes the mcp__<name>__<tool> prefix), url, transport (streamable-http | sse; NO stdio per ADR-018), encrypted authHeaders (_secret-json-column envelope), enabled, source ('manual' now, 'package' reserved), lastConnectedAt/lastError status stamps.
  • Per-agent bindings (agent_mcp_server_bindings, per plan §2.5 naming): 'tenant' rows inherit a connection to ALL of the user's agents (created enabled=true alongside the connection), 'agent' rows override per agent (narrow-only semantics like tool grants; delete the row to revert to inheritance). Unique (connectionId, targetType, targetId).
  • packages/agent/src/mcp/ (new module, exported as @ever-works/agent/mcp):
    • McpClientService — official @modelcontextprotocol/sdk (^1.27.1, same as apps/mcp) behind the MCP_CLIENT_FACTORY seam (structural interfaces, lazy import(); tests inject fakes — no SDK, no network). listTools with a 60s TTL cache per connection; callTool with 30s timeout + 100KB serialized-result cap; errors classified into short header-free messages and stamped on the row. Auth header values never reach logs or errors.
    • McpConnectionsService — CRUD + masking (responses carry authHeaderNames only), name/URL/header validation (lexical SSRF guard isSafeWebhookUrl on URLs), test endpoint, per-agent binding state + effective-connection resolution. Cross-user access → 404.
    • McpToolSource — bound connections → AgentToolDescriptors named mcp__<server>__<tool> (schema passthrough, description prefixed [<server>], name/description sanitized, length caps). Dead server → zero tools + WARN, never a failed run. Gated on permissions.canCallExternalTools (outbound-call risk class).
  • Tool funnel integration: new AGENT_MCP_TOOL_SOURCE token (packages/agent/src/agents/agent-mcp-tool-source.ts), injected @Optional() APPENDED LAST into AgentToolService; consumed inside resolveGrantedTools (async) BEFORE the grant partition, so mcp__* names flow through partitionToolsByGrant and the run-level funnel for free. Built-in name collisions are dropped with a WARN, and each accepted descriptor is wrapped in the same withCredentialInterpolation the built-ins get — MCP servers are the outbound calls {{cred.key}} exists for, so an unwrapped descriptor would forward the literal template text to a third party and scrub nothing on the way back. Bound in the api-side @Global AgentsModule (useExisting: McpToolSource) + pin-spec updated.
  • Activity: additive ActivityActionType members mcp_connection_created/updated/deleted/tested, mcp_binding_updated (varchar storage — no migration). Emitted from McpConnectionsService.
  • Run-log: no change needed — MCP invocations ride the existing tool-loop logging; toolName carries the mcp__ prefix end-to-end.

Data model / migration​

apps/api/src/migrations/1786840000000-CreateMcpServerConnections.ts — portable Table API (postgres + better-sqlite3), idempotent guards, both tables + unique indexes, FKs to users (+ bindings → connections CASCADE). Entities registered in entities/index.ts, _entities-inventory.ts, and _entity-names.ts (both halves of the two-step registration).

Endpoints​

  • GET/POST /api/mcp-connections, GET/PATCH/DELETE /api/mcp-connections/:id, POST /api/mcp-connections/:id/test (returns {ok, toolCount, tools[]}).
  • GET /api/agents/:agentId/mcp-servers (connections + effective state incl. inheritedFromTenant), PUT /api/agents/:agentId/mcp-servers/:connectionId {enabled}, DELETE …/:connectionId (revert to inherit).
  • Module apps/api/src/mcp-connections/, registered in api.module.ts.

UI routes​

  • Settings → Connections (/settings/connections): list + add form (name/url/transport/one auth header), enable/disable switch, Test (shows tool count + first names), delete, last-status line. New tab in settings-layout-client.tsx (Plug icon).
  • Agent detail → MCP Servers tab (/agents/[id]/mcp-servers): all connections with effective toggles, Inherited/Override/Connection-disabled badges, Revert-to-inherit. Tab added in AgentDetailTabs.tsx; route constant DASHBOARD_AGENT_MCP_SERVERS.
  • i18n keys added to ALL 21 apps/web/messages/*.json (English values copied per convention).

Tests​

  • packages/agent (Jest): cd packages/agent && npx jest --testPathPattern='(src/mcp/|agent-tool-mcp-funnel)' — 37 tests. Client cache TTL/timeout/size-cap/error-classification/secret masking; binding resolution matrix (tenant-inherit / agent-disable / agent-only / disabled-connection); descriptor naming + sanitization + executor proxy + dead-server isolation; connections CRUD + masking + SSRF + cross-user 404 + test endpoint + per-agent state. agents/__tests__/agent-tool-mcp-funnel.spec.ts is the FUNNEL half: no-source parity, descriptors appended, built-in collision dropped, grant-matrix deny of an mcp__* name (proves the append happens before partitionToolsByGrant), a throwing source never failing run assembly, and {{cred.key}} interpolation reaching MCP args. Updated pin spec: entities/__tests__/activity-log.types.spec.ts (+5 MCP action types).
  • apps/api (Jest): cd apps/api && npx jest --testPathPattern='(agents.module|mcp-connections|CreateMcpServerConnections)' — pin spec asserts McpModule import + AGENT_MCP_TOOL_SOURCE binding/export; controller delegation/auth-scoping pins; the migration spec runs both tables' up/down on better-sqlite3 (columns, indexes, idempotency, both unique constraints).

Verification run on this branch​

  • cd packages/agent && npx tsc -p tsconfig.types.json --noEmit — clean.
  • npx turbo build --filter=@ever-works/agent — passes.
  • cd packages/agent && npx jest --testPathPattern='(src/mcp/|activity-log.types|portable-date-columns|agent-tool)' — 12 suites / 219 tests green (regression sweep over every agent-tool* suite, since the funnel changed).
  • cd apps/api && npx jest — 254 suites / 4030 tests green.
  • cd apps/api && npx tsc --noEmit — 4 PRE-EXISTING errors unrelated to this branch (@ever-works/k8s-plugin missing in deploy.e2e.spec.ts; three @src/* alias misses in packages/agent sources reached through a relative path). None involve MCP files; not fixed here.
  • cd apps/web && npx tsc --noEmit — clean.
  • cd apps/web && npx eslint <the six new web files> — clean. (packages/agent and apps/api have no lint script and no eslint config — pnpm lint only reaches apps/web.)
  • npx prettier --check over every touched file — clean (root config resolves to .prettierrc: 4-space, 100 cols, trailing commas).
  • i18n audit: all 10 new key paths present in all 21 apps/web/messages/*.json files, English values copied.

Decisions & divergences from the brief​

  • Bindings are consumed in resolveGrantedTools (async), not the sync resolveAllowedTools: descriptor assembly needs I/O and the run path (agent-run.service.ts:1480-1488) already prefers the async method. Sync-only callers (no grant enforcer path) see no MCP tools — acceptable: production runs go through resolveGrantedTools.
  • Creating a connection auto-creates the enabled tenant binding so a manual connection is usable immediately (the brief's "global to the workspace"), narrowable per agent. No separate tenant-binding management UI in v1.
  • MCP tools are gated on canCallExternalTools — same outbound-network risk class as searchWeb/screenshot/extractContent/sendEmail. Not in the brief, consistent with the funnel's posture.
  • skill-binding-style @ManyToOne to the connection/user rows (the no-@ManyToOne rule applies to SCOPE entities only; tenant/org stay raw uuid columns).
  • lastConnectedAt is @PortableDateColumn (NOT raw type: 'timestamp'): the e2e/CI driver is better-sqlite3, which has no timestamp type — a raw timestamp fails TypeORM metadata validation at boot. The repo-wide entities/__tests__/portable-date-columns.spec.ts guard SCANS entity sources, so it needed no edit; it simply had to pass, which is what forced the column type. The MIGRATION still uses timestamp (house pattern — migrations run on Postgres only); TypeORM maps type: Date to timestamp there, so entity and schema agree.

Known follow-ups​

  • Account-transfer whitelist entries for the two new tables (export/import of connections as references, masked secrets) — plan §2.5's 4-place whitelist work was out of budget here; new tables simply do not transfer.
  • Playwright e2e for the two new pages (binding CRUD, authz matrix).
  • No tenant-binding management UI: POST /api/mcp-connections creates the enabled 'tenant' row and nothing exposes it afterwards, so "disable everywhere" is done by disabling the CONNECTION. Fine for v1 (they are equivalent while manual is the only source), but the tenant row needs its own control once 'package' connections exist.
  • Domain chat tools for connections/bindings (the in-code DoD at agent-tool.service.ts:317-319) — new entities have no chat-tool source yet.
  • Redirect policy (AP-15 cross-origin header rules) relies on the SDK's fetch defaults; the explicit no-forward implementation is deferred to the package slice (T25).
  • Per-run connection pooling (connect-per-operation today; fine at v1 scale).