feat: import OpenAPI and GraphQL specs as MCP tools

[daltoniam]

Summary

• Adds a "bring your own spec" feature: paste an OpenAPI 3.x document or a GraphQL introspection result and its operations become searchable, executable MCP tools — no adapter code required.
• New management page in the web UI (list / add / remove) with live status, plus live reload so imported tools appear without restarting the server.

How it works

• New specimport/ package parses a spec into a runtime Integration (Load), and a Syncer reconciles the registry against config on startup and on every config change — it only ever touches integrations it registered, never built-ins.
• Imports are stored in config under spec_imports and injected server-side credentials never reach the model.

Fixes bundled in

Two bugs discovered during live testing that made imported tools invisible to search/execute even though they showed as active in the UI:

1. Config loading silently dropped imports on every read (they weren't carried through the defaults merge).
2. The enabled-integration lookup used by search indexing and execute resolution didn't include imports.

Verification

Tested end-to-end against the live Swagger Petstore: search returns the imported tools and execute performs real upstream calls (valid IDs return data, invalid IDs return the upstream 404). Full CI passes (build, vet, race tests, lint).

💘 Generated with Crush

[acmacalister]

CI/status checks all passing (build, test, lint, security). Clean PR — follows project conventions, comprehensive test coverage, no blocking issues found. LGTM.

[acmacalister]

Solid feature — clean Syncer boundary, thorough parser/execute coverage, and the EnabledIntegrations/config-merge fixes are the right root-cause fixes. CI is green across build/test/lint/security. A few concrete gaps inline around auth_scheme, name collisions/dedupe, and dead IsWrite wiring.

[acmacalister]

The form copy says leave auth_scheme empty to send a raw key, but we only persist non-empty values here. An empty field never lands in the credentials map, so Configure takes the !hasScheme path and still defaults to Bearer.

Either stop advertising the empty-scheme case in the UI, or persist an explicit sentinel (e.g. always write the key when api_key is set, including auth_scheme: "") so hasScheme is true:

if apiKey := strings.TrimSpace(r.FormValue("api_key")); apiKey != "" {
    creds["api_key"] = apiKey
    // Always record scheme when a key is present so empty means "raw".
    creds["auth_scheme"] = strings.TrimSpace(r.FormValue("auth_scheme"))
    ...
}

[acmacalister]

Load succeeding is necessary but not sufficient — registration can still fail when the sanitized name collides with a built-in (github, slack, web, …). Those are always in the registry at startup, so Sync hits Register → already registered, the entry stays in config as Error, and if that built-in is also enabled, EnabledIntegrations emits the name twice and search indexes the same tools twice.

Worth rejecting reserved/colliding names up front (against Registry.Names() or a fixed built-in set) before persisting, with a clear flash like name \"github\" is reserved by a built-in integration.

[acmacalister]

This appends every enabled import’s sanitized name with no dedupe. Two config entries that collapse to the same identifier (Demo API + demo-api), or an import whose name matches an enabled built-in, make EnabledIntegrations return duplicates. Downstream search/index paths iterate this list as-is, so tools get indexed twice.

A small seen-set (or skipping names already present from Integrations) would close that:

seen := make(map[string]struct{}, len(names))
for _, n := range names { seen[n] = struct{}{} }
for _, si := range m.cfg.SpecImports {
    if !si.Enabled { continue }
    n := specimport.SanitizedName(si.Name)
    if _, ok := seen[n]; ok { continue }
    seen[n] = struct{}{}
    names = append(names, n)
}

[acmacalister]

IsWrite is implemented and tested, but nothing outside this package calls it — there’s no WriteAware (or similar) port on mcp.Integration, and the server never type-asserts it. Right now the only write signal that reaches the model is the [write] description prefix.

Either wire this through a real optional interface the policy/search layer consults, or drop the method (and the “policy layer uses this” comments) so we don’t imply write gating that isn’t actually enforced.

[acmacalister]

Get() hands back the live config pointer, and on a hit this assigns into that shared slice before SetSpecImports runs. Concurrent readers of SpecImports (or a racing save) can observe a half-updated entry with no lock held.

Safer to copy first, then mutate the copy:

out := make([]mcp.SpecImportConfig, len(list))
copy(out, list)
// replace or append on out
return out

(removeSpecImport already avoids this via list[:0:0].)

[acmacalister]

CI is green across build/test/lint/security. Prior feedback looks addressed (auth_scheme sentinel, reserved names, EnabledIntegrations dedupe, upsert copy, IsWrite removal). Two small follow-ups inline — credential wipe on re-save, and docs that still imply write-approval gating.

[acmacalister]

Re-saving an existing import with a blank API key wipes credentials. entry starts with a nil Credentials map, and we only assign when the form sends something — so an upsert of the same name replaces the stored key with nothing.

Worth merging with the previous entry when the form leaves auth fields empty, e.g. keep list[i].Credentials unless the form actually supplied new ones. Password inputs being empty on re-submit is the usual case here.

[acmacalister]

After dropping IsWrite, this package doc (and a few sibling comments in openapi.go / integration.go) still says writes are "marked destructive so the policy layer can gate them." In-tree there’s no write-approval / destructive-hint path — the only signal is the [write] description prefix.

Might be worth toning these down to match what’s actually enforced, so we don’t imply gating that isn’t wired yet.