Salesforce Models API setup for an AI coding agent
The Salesforce Models API (https://api.salesforce.com/ai/gpt/v1) is authenticated with a signed OrgJWT (obtained via clientcredentials with the sfapapi scope — see scripts/get-orgjwt.sh; no proxy). That auth and the base URL are the same for any agent. How each agent then talks to the endpoint is agent-specific: Anthropic clients (Claude Code and the Claude Agent SDK) route through Bedrock mode (the env vars in Step 3), whereas other agents (e.g. Codex) use their own client config against the same endpoint and token — Bedrock mode does not apply to them.
The steps below are the Claude Code / Claude Agent SDK reference implementation (Bedrock mode + a JSON settings file + an API-key helper). For a non-Bedrock agent, reuse the OrgJWT auth (Step 1) and the base URL, and apply the equivalent client settings in that agent's own config location instead of the Bedrock env vars.
Bundled scripts are in scripts/. Path placeholders below: <SKILL> = the absolute path to this skill's own directory (the folder containing this SKILL.md; resolve it from the skill path in context). <ABS> = the absolute path to the user's project root. Always emit fully resolved absolute paths — the API-key helper runs from an undefined working directory, so relative paths break it.
Prerequisite
A connected app in the org with the sfapapi OAuth scope and the clientcredentials flow enabled (consumer key/secret + a run-as user). Setup steps: https://developer.salesforce.com/docs/ai/agentforce/guide/access-models-api-with-rest.html curl + jq installed.
Inputs to collect
SFINSTANCEURL — org My Domain, e.g. https://acme.my.salesforce.com
SFCLIENTID, SFCLIENTSECRET — connected-app consumer key/secret
- Models API base URL:
https://api.salesforce.com/ai/gpt/v1
- Model: a fully qualified
sfdc_ai__… name, e.g.
sfdc_ai__DefaultBedrockAnthropicClaude46Sonnet (full list: https://developer.salesforce.com/docs/ai/agentforce/guide/supported-models.html)
- Scope: project (
<cwd>/.claude/settings.json, default) or user (~/.claude/settings.json) — reference-agent settings paths
- Headers —
<FEAT> = x-client-feature-id (default ai-platform-models-connected-app),
<APP> = x-sfdc-app-context (default EinsteinGPT). Used in the Step 2 verify curl and in ANTHROPICCUSTOMHEADERS.
Steps (reference implementation)
Concrete values for a JSON-settings + API-key-helper agent. Reuse the OrgJWT auth, verify curl, and base URL verbatim for any agent; adapt the settings-file location and env-var wiring to the target agent.
- Write
<project>/.claude/.orgjwt.env (chmod 600), gitignore it:
``ini SFINSTANCEURL="..." SFCLIENTID="..." SFCLIENTSECRET="..." ``
- Verify — must return
200 before writing settings:
``bash TOKEN=$(bash <SKILL>/scripts/get-orgjwt.sh <ABS>/.claude/.orgjwt.env) curl -s -o /dev/null -w '%{httpcode}\n' \ <MODELSAPIURL>/model/<MODEL>/invoke-with-response-stream \ -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \ -H 'x-client-feature-id: <FEAT>' -H 'x-sfdc-app-context: <APP>' \ --data '{"anthropicversion":"bedrock-2023-05-31","max_tokens":16,"messages":[{"role":"user","content":"hi"}]}' ``
- Write
.claude/settings.json (merge into existing; keep other keys):
``json { "apiKeyHelper": "bash <SKILL>/scripts/get-orgjwt.sh <ABS>/.claude/.orgjwt.env", "model": "<MODEL>", "env": { "ANTHROPICAUTHTOKEN": "", "CLAUDECODEUSEBEDROCK": "1", "CLAUDECODESKIPBEDROCKAUTH": "1", "ANTHROPICBEDROCKBASEURL": "<MODELSAPIURL>", "ANTHROPICSMALLFASTMODEL": "<MODEL>", "ANTHROPICDEFAULTMODEL": "<MODEL>", "ANTHROPICCUSTOM_HEADERS": "x-client-feature-id: <FEAT>\nx-sfdc-app-context: <APP>" } } ` Use absolute paths in apiKeyHelper. (<FEAT> / <APP>` defaults are in "Inputs to collect" above.)
- Tell the admin to fully restart the agent (
claude for the reference agent) —
settings and the API-key helper load at startup only.
Capturing as a runbook (when asked to document, not apply)
If the user wants the setup written up for review instead of applied to their machine (e.g. "save it as a Markdown runbook"), write all of the above into the requested file (e.g. models-api-setup-runbook.md), in order and self-contained: the exact .orgjwt.env contents, the chmod 600 + gitignore note, the verification curl (with the "must be 200 before writing settings" note), the full settings.json block with every key from Step 3, and the final "fully restart claude" step. Don't omit any of the nine settings.json keys.
Verify before finishing
Must be exact (each prevents a specific failure)
"ANTHROPICAUTHTOKEN": "" — clears any global token that would otherwise
outrank apiKeyHelper (precedence: ANTHROPICAUTHTOKEN > ANTHROPICAPIKEY > apiKeyHelper). Without it → wrong/old bearer → 401/404.
CLAUDECODEUSE_BEDROCK=1 — activates the Bedrock API client; without it
Claude Code uses the standard Anthropic API protocol and ignores ANTHROPICBEDROCKBASE_URL entirely, so every call bypasses the Models API.
CLAUDECODESKIPBEDROCKAUTH=1 — else Claude Code overwrites Authorization
with AWS SigV4 and the OrgJWT never lands.
apiKeyHelper must be invoked as bash <path> <credsfile> (avoids exit-126).
- Model must be a fully qualified
sfdc_ai__… name (see supported models).
- Auth is the OrgJWT from
client_credentials (a signed JWT, 2 dots, scope
sfapapi) — NOT sf org display (unsigned session token → 404). sf CLI has no clientcredentials command; the helper calls /services/oauth2/token.
- Only
ANTHROPICBEDROCKBASE_URL routes; no tenant-id header needed.
Diagnose
| Error |
Meaning |
Check first |
401 |
Token is not a valid OrgJWT |
Connected App sfapapi scope, clientcredentials flow enabled, consumer key/secret in .orgjwt.env; ANTHROPICAUTHTOKEN not cleared to "" |
404 |
Token valid but model/env/org not routable |
Fully qualified sfdc_ai__… model alias, ANTHROPICBEDROCKBASEURL exactly https://api.salesforce.com/ai/gpt/v1, org entitled for the Models API, ANTHROPICAUTH_TOKEN cleared |
model not available |
Non-alias model id |
Replace with a fully qualified sfdc_ai__… alias (see supported models) |