sf-connected-apps: Salesforce Connected Apps & External Client Apps
Use this skill when the user needs OAuth app configuration in Salesforce: Connected Apps, External Client Apps (ECAs), JWT bearer setup, PKCE decisions, scope design, or migration from older Connected App patterns to newer ECA patterns.
Brite Context
Brite's integration auth landscape sits mid-migration. Spring '26 disabled org-wide creation of classic Connected Apps (SF Support re-enable required — treated as temporary migration runway), so every net-new OAuth app must be an ExternalClientApplication. Current inventory:
Marketing_Claude_MCP— Brite's first pure ECA, provisioned post-Spring-'26 per BC-5579. Plugin-side MCP runtime access. JWT bearer. ScopesApi+RefreshToken. Zero-persistence refresh-token policy — as a pure ECA the field isrefreshTokenPolicyTypeonExtlClntAppOauthConfigurablePolicies, not ConnectedApp'srefreshTokenPolicy. Private key in the Engineering Bitwarden collection.Outbound_Sales_Ops— live runtime integration writing CF-owned fields on Lead/Contact. Legacy ConnectedApp auto-wrapped with an ECA settings file during the Spring '26 migration — NOT a pure ECA. Parent.connectedApp-meta.xml+ wrapper.eca-meta.xml. JWT bearer.isAdminApproved=true,refreshTokenPolicy=zero,ipRelaxation=ENFORCE.CI_Deploy— ECA committed but inactive ("compliance-posture-only"). CI does NOT use this app. GitHub Actions deploys authenticate viaSFDX_AUTH_URLunder Salesforce's built-inPlatformCLIConnected App.OutboundSync— separate sync path.
Credential home: Engineering Bitwarden collection. Naming convention and binding rule in Convention 7.
See also: brite-salesforce/CLAUDE.md §External Client Apps, docs/research/salesforce-mcp-findings.md Q3 + Appendix B.1, and memory gotcha_spring26_ca_blocked.
Brite ECA Conventions
These rules bind any sf-connected-apps work done in a Brite repo. Violations block PR review.
Use
ExternalClientApplicationfor new OAuth apps. ClassicConnectedAppcreation is disabled org-wide since Spring '26. Even for a simple single-org app, use the ECA metadata type. The upstream "First Decision" table below still lists Connected App as preferred for some cases — treat those rows as upstream reference only; at Brite the answer is always ECA.JWT-via-SFDX-CLI flow requires BOTH
ApiandRefreshTokenscopes. Thesf org login jwtcommand persists auth records in~/.sfdx/and needsRefreshTokenfor that persistence — drop it and JWT exchange fails with "refresh_token scope is required." Pure JWT exchange at a non-CLI runtime (e.g. a Python MCP calling the token endpoint directly) does NOT requireRefreshToken— apply the scope requirement where SFDX handles the handshake, not universally. Source:docs/research/salesforce-mcp-findings.mdQ3 lines 50-52.Default to a zero refresh-token policy, and mind which type you are editing. On a ConnectedApp the field is
<refreshTokenPolicy>inside<oauthPolicy>, and the enum is lowercase —infinite | zero | specific_lifetime; uppercaseZEROwill not validate. Seeassets/connected-app-jwt.xml. On a pure ECA the field isrefreshTokenPolicyTypeonExtlClntAppOauthConfigurablePolicies(assets/eca-oauth-policies.xml) — a different type with a different enum, so do not copy the ConnectedApp value across. The ECA zero-persistence value is not yet recorded here.Marketing_Claude_MCPalready runs one, so retrieve its.ecaOauthPlcyand write the value into this rule and into the asset. Until that lands, this convention is enforceable for ConnectedApps and advisory for ECAs — do not fail a review on a value the skill cannot supply. Belt-and-suspenders against long-lived refresh tokens even when clients mis-request them; default for new ECAs unless a flow documents a specific persistence need in its PR.Exclude
ExtlClntAppOauthSettingsvia.forceignorefor sandbox deploys. The file embeds an org-specificoauthLink(OrgId:ConsumerRecordId) that does not resolve cross-org. Excluded alongside theExternalClientApplicationtype for sandbox work; temporarily comment the exclusions out before running a production deploy, then re-enable. The canonical toggle procedure lives in sf-deploy.JWT-from-ECA breaks
sf org create scratch. Upstream bugs:forcedotcom/cli#3025,forcedotcom/cli#3482. Scratch-org creation rejects JWT sessions authenticated against ECAs. Workaround: authenticate viaSFDX_AUTH_URLthrough the CLI's built-inPlatformCLIConnected App for scratch-org provisioning. Runtime MCP calls are unaffected.CI auth is
SFDX_AUTH_URLvia built-inPlatformCLI— notCI_Deploy. TheCI_DeployECA in source is compliance-posture-only (committed but inactive). GitHub Actions deploys authenticate via anSFDX_AUTH_URLsecret loaded into the CLI's built-inPlatformCLIConnected App. Runtime MCP access goes through JWT againstMarketing_Claude_MCP. Don't repurposeCI_Deploywithout explicit ADR approval — the blast-radius story depends on the CI path and runtime path staying independent.All ECA secrets live in the Engineering Bitwarden collection, never in source. JWT private keys and consumer secrets are stored with item names
<ECAName> — JWT private keyand<ECAName> — consumer secret. Per-user distribution is out-of-band. Never commit, never echo to logs, never put in env templates with real values.
When This Skill Owns the Task
Use sf-connected-apps when the work involves:
.connectedApp-meta.xmlor.eca-meta.xmlfiles- OAuth flow selection and callback / scope setup
- JWT bearer auth, device flow, client credentials, or auth-code decisions
- Connected App vs External Client App architecture choices
- consumer-key / secret / certificate handling strategy
Delegate elsewhere when the user is:
- configuring Named Credentials or runtime callouts → sf-integration
- analyzing access / permission policy assignments → sf-permissions
- writing Apex token-handling code → sf-apex
- deploying metadata to orgs → sf-deploy
First Decision: Connected App or External Client App
At Brite: always ECA. See § Brite ECA Conventions rule 1. The table and default guidance below are upstream reference only — rows recommending Connected App do not apply in Brite repos because Spring '26 blocks new Connected App creation org-wide.
| If the need is... | Prefer |
|---|---|
| simple single-org OAuth app | Connected App |
| new development with better secret handling | External Client App |
| multi-org / packaging / stronger operational controls | External Client App |
| straightforward legacy compatibility | Connected App |
Default guidance:
- choose ECA for new regulated, packageable, or automation-heavy solutions
- choose Connected App when simplicity and legacy compatibility matter more
- Spring ’26 note: creation of new Connected Apps is disabled org-wide (SF Support re-enable required, treated as temporary migration runway). At Brite, always use ECA — see Brite ECA Conventions rule 1.
Required Context to Gather First
Ask for or infer:
- app type: Connected App or ECA
- OAuth flow: auth code, PKCE, JWT bearer, device, client credentials
- client type: confidential vs public
- callback URLs / redirect surfaces
- required scopes
- distribution model: local org only vs packageable / multi-org
- whether certificates or secret rotation are required
Recommended Workflow
1. Choose the app model
Decide whether a Connected App or ECA is the better long-term fit.
2. Choose the OAuth flow
| Use case | Default flow |
|---|---|
| backend web app | Authorization Code |
| SPA / mobile / public client | Authorization Code + PKCE |
| server-to-server / CI/CD | JWT Bearer |
| device / CLI auth | Device Flow |
| service account style app | Client Credentials (typically ECA) |
3. Start from the right template
Use the provided assets instead of building from scratch:
assets/connected-app-basic.xmlassets/connected-app-oauth.xmlassets/connected-app-jwt.xmlassets/external-client-app.xmlassets/eca-global-oauth.xmlassets/eca-oauth-settings.xmlassets/eca-policies.xml— app enablement + start page ONLYassets/eca-oauth-policies.xml— permitted users, IP relaxation, refresh-token lifetime, session level
The two policy types are near-identically named and carry disjoint fields. ExtlClntAppConfigurablePolicies (.ecaPlcy) holds externalClientApplication, isEnabled, isOauthPluginEnabled, label, startPage. Everything that governs whether a credential actually works lives on ExtlClntAppOauthConfigurablePolicies (.ecaOauthPlcy) — and its field names are not ConnectedApp's: ipRelaxationPolicyType (not ipRelaxation), refreshTokenPolicyType (not refreshTokenPolicy). Putting OAuth policy on the .ecaPlcy type produces a file that will not deploy.
If you need source-controlled ECA OAuth security metadata, retrieve it from an org first and treat the retrieved file as the schema source of truth:
sf project retrieve start --metadata ExtlClntAppOauthSecuritySettings:<AppName> --target-org <alias>
4. Apply security hardening
Favor:
- least-privilege scopes
- explicit callback URLs
- PKCE for public clients
- certificate-based auth where appropriate
- rotation-ready secret / key handling
- IP restrictions when realistic and maintainable
5. Validate deployment readiness
Before handoff, confirm:
- metadata file naming is correct
- scopes are justified
- callback and auth model match the real client type
- secrets are not embedded in source
High-Signal Security Rules
Avoid these anti-patterns:
| Anti-pattern | Why it fails |
|---|---|
| wildcard / overly broad callback URLs | token interception risk |
Full scope by default |
unnecessary privilege |
| PKCE disabled for public clients | code interception risk |
| consumer secret committed to source | credential exposure |
| no rotation / cert strategy for automation | brittle long-term ops |
Default fix direction:
- narrow scopes
- constrain callbacks
- enable PKCE for public clients
- keep secrets outside version control
- use JWT certificates or controlled secret storage where appropriate
ECA Pre-Authorization and JWT Bearer
These are the settings that decide whether a JWT credential works at all. All measured against brite-dev-kells, 2026-07-28.
Permitted Users must be AdminApprovedPreAuthorized
JWT has no interactive consent step, so a self-authorize policy can never be satisfied. The failure is user is not admin approved to access this app. Set permittedUsersPolicyType accordingly on ExtlClntAppOauthConfigurablePolicies.
The pre-auth permission-set picker lists custom permsets only
As of 2026-07-28 it listed IsCustom = true permission sets only. A standard licensed permset — SalesEngagementBasicUser, Type = Standard — did not appear, no matter how correct it looked. That leaves two options: a profile grant, which authorizes everyone on that profile, or a purpose-built custom handle permset. Prefer the handle permset. (Picker behaviour is UI-version-sensitive; re-check before concluding a permset is unselectable today.)
The certificate upload only appears after JWT Bearer Flow is enabled
Tick Enable JWT Bearer Flow first; the certificate field renders only then. There is no "Use digital signatures" checkbox as on Connected Apps — looking for one is a genuine dead end.
This is a Setup-UI path. The metadata type carrying the flow-enablement flags was not captured in the 2026-07-28 audit — ExtlClntAppOauthConfigurablePolicies does not carry it, and ExtlClntAppOauthSecuritySettings is the likely home but unverified. Retrieve before assuming a metadata-only route exists.
High Assurance session level breaks JWT outright
A server-to-server session cannot satisfy MFA. If requiredSessionLevel demands High Assurance the exchange fails in a way that looks nothing like an auth problem. STANDARD means High Assurance off.
Wrong audience fails as invalid_grant: invalid assertion
The per-environment audience mapping already lives in references/oauth-flows-reference.md (sandbox is https://test.salesforce.com). What is worth adding is the symptom: using login.salesforce.com against a sandbox returns invalid_grant: invalid assertion, which reads like a certificate fault and sends you to the wrong place entirely.
Verifying ECA Auth — Two Techniques Worth Keeping
sf org login jwt exits 0 on a FAILED JWT
Observed directly on 2026-07-28: a login returning user is not admin approved to access this app still exited 0. The same false-success is recorded for sf hardis auth login in the brite-salesforce repo; it applies to the raw CLI command too. Never treat exit status as proof. Always confirm with a real query against the org.
Falsify a pre-auth grant without touching the live credential
To test whether a pre-authorization grant is load-bearing, run JWT as a second user on the same profile who lacks the permset. The holder authorizes and the non-holder fails, so the permset is the only variable — and you never deselect a grant on a working credential.
This settled a real dispute. docs/runbooks/ci-deploy-user-provisioning.md (BC-13242) records permset-only pre-auth failing in prod for ci@, which turned out not to generalize.
An Empty Metadata Retrieve Is Not Evidence of a Platform Limitation
Two wrong conclusions were drawn in a row from the same empty retrieve — first "the type is org-local and Setup-UI-only", then "wildcards don't enumerate these types". Both false.
The real cause: ExtlClntAppGlobalOauthSettings and the policy records are created when OAuth is enabled on the app, and the retrieve ran before that. Wildcards and type-level retrieves both work fine.
The lesson generalizes past ECAs. Before concluding a metadata type cannot be captured, run:
sf org list metadata --metadata-type <TypeName> --target-org <alias>
That distinguishes "does not exist yet" from "cannot be captured" in one call.
Security Note: .ecaGlblOauth Carries the Consumer Key in Plaintext
ExtlClntAppGlobalOauthSettings stores the consumerKey in plaintext beside the base64 certificate. Those two fields appear only on a retrieved file — the authoring template at assets/eca-global-oauth.xml carries neither, which is why the two look inconsistent.
This skill is what tells people to retrieve these types, so treat a retrieved .ecaGlblOauth as sensitive: keep it out of commits, and check the consuming repo's .forceignore before retrieving into a working tree. Same mechanism as Convention 4 above, applied to a different type.
Metadata Notes That Matter
Connected App
Usually lives under:
force-app/main/default/connectedApps/
External Client App
Current source-supported ECA metadata uses multiple top-level source directories, not a single externalClientApps/ folder:
force-app/main/default/externalClientApps/→ExternalClientApplication(.eca-meta.xml)force-app/main/default/extlClntAppGlobalOauthSets/→ExtlClntAppGlobalOauthSettings(.ecaGlblOauth-meta.xml)force-app/main/default/extlClntAppOauthSettings/→ExtlClntAppOauthSettings(.ecaOauth-meta.xml)force-app/main/default/extlClntAppOauthSecuritySettings/→ExtlClntAppOauthSecuritySettings(.ecaOauthSecurity-meta.xml)force-app/main/default/extlClntAppOauthPolicies/→ExtlClntAppOauthConfigurablePolicies(.ecaOauthPlcy-meta.xml)force-app/main/default/extlClntAppPolicies/→ExtlClntAppConfigurablePolicies(.ecaPlcy-meta.xml)
Important file-name gotchas:
- the global OAuth suffix is
.ecaGlblOauth, not.ecaGlobalOauth - the general policy suffix is
.ecaPlcy, not.ecaPolicy - use
.ecaOauthSecurityforExtlClntAppOauthSecuritySettings
Output Format
When finishing, report in this order:
- App type chosen
- OAuth flow chosen
- Files created or updated
- Security decisions
- Next deployment / testing step
Suggested shape:
App: <name>
Type: Connected App | External Client App
Flow: <oauth flow>
Files: <paths>
Security: <scopes, PKCE, certs, secrets, IP policy>
Next step: <deploy, retrieve consumer key, or test auth flow>
Cross-Skill Integration
| Need | Delegate to | Reason |
|---|---|---|
| Named Credential / callout runtime config | sf-integration | runtime integration setup |
| deploy app metadata | sf-deploy | org validation and deployment |
| Apex token or refresh handling | sf-apex | implementation logic |
| permission review after deployment | sf-permissions | access governance |
Reference Map
Start here
- references/oauth-flows-reference.md
- references/security-checklist.md
- references/testing-validation-guide.md
Migration / examples
- references/migration-guide.md
- references/example-usage.md
- assets/
Score Guide
| Score | Meaning |
|---|---|
| 80+ | production-ready OAuth app config |
| 54–79 | workable but needs hardening review |
| < 54 | block deployment until fixed |