Skip to main content
Provider files define the execution contract between Promptbeat and your agent or LLM. They specify the provider type, model, and any configuration the target runtime needs. Because provider files are separate from your scenario and generation config, you can regenerate probes once and re-evaluate them against multiple targets without touching the scenario logic. This separation is one of the core design principles that makes Promptbeat comparisons reproducible.

​
What a provider file contains

Every provider file contains a providers list. Each entry in the list has three parts:
  • id — the provider type string that selects the execution backend, such as openai:codex-sdk, openai:gpt-4o, anthropic:claude-3-5-sonnet-latest, or http.
  • label — a human-readable name shown in reports and artifact filenames.
  • config — a block of provider-specific settings such as model, working_dir, sandbox_mode, url, headers, and body.
providers.codex-sdk.yaml
providers:
  - id: openai:codex-sdk
    label: Codex SDK via Promptfoo
    config:
      model: openai/gpt-5.4
      working_dir: agent-workspace
      sandbox_mode: read-only
      approval_policy: never
      skip_git_repo_check: true
      enable_streaming: false
      deep_tracing: false
      inherit_process_env: true
providers.http.yaml
providers:
  - id: http
    label: Support agent HTTP endpoint
    config:
      url: http://localhost:8088/eval/run
      method: POST
      headers:
        content-type: application/json
        authorization: Bearer {{env.AGENT_EVAL_TOKEN}}
      body:
        prompt: '{{prompt}}'
        scenario_id: '{{scenarioId}}'
        tenant_id: tenant-a
        user_id: user-123
      transformResponse: json.answer

​
Path handling

Pass your provider file to Promptbeat using the --provider-file flag on generate and eval. Promptbeat resolves working_dir and cwd values inside the config block relative to the provider file’s own location — not relative to your shell’s current directory. This keeps paths stable when generated configs are written to an artifacts/ subdirectory. 
uv run promptbeat generate \
  --config promptbeat.yaml \
  --provider-file providers.codex-sdk.yaml \
  --generator-provider openai:openai/gpt-5.5 \
  --count 20 \
  --output-dir artifacts/codex/generate

uv run promptbeat eval \
  --config artifacts/codex/generate/generated_redteam.yaml \
  --provider-file providers.codex-sdk.yaml \
  --output-dir artifacts/codex/eval

​
Credential handling

Provider YAML files are logged in Promptbeat artifacts. Any credential written directly into a provider file will appear in your evaluation output. Never commit API keys to a provider file.
Use environment variables for all credentials. Reference them in provider YAML with the {{env.VAR_NAME}} syntax — Promptbeat substitutes the value at runtime and never writes the literal key into generated files.
Environment variableUsed by
OPENAI_API_KEYOpenAI LLM providers and the Codex SDK provider
ANTHROPIC_API_KEYAnthropic and Claude model providers
GOOGLE_API_KEYGemini model providers
AGENT_EVAL_TOKENHTTP agent providers (any name you choose)
OPENCLAW_API_KEYOpenClaw gateway provider
Set variables in your shell before running Promptbeat: 
export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export GOOGLE_API_KEY="AIza..."
For the Codex SDK provider specifically, prefer an explicit cli_env allowlist over inherit_process_env: true in production. The allowlist passes only the variables Codex needs, rather than the entire invoking shell environment: 
    config:
      inherit_process_env: false
      cli_env:
        OPENAI_API_KEY: '{{env.OPENAI_API_KEY}}'
        CODEX_HOME: ./sample-codex-home

​
Multiple providers

Run a multi-provider comparison by listing more than one provider in a single file or by passing --provider-file more than once. Promptbeat routes each generated probe to all listed providers and writes a combined report that shows pass rates side by side.
providers.comparison.yaml
providers:
  - id: openai:codex-sdk
    label: Codex gpt-5.4
    config:
      model: openai/gpt-5.4
      working_dir: agent-workspace
      sandbox_mode: read-only
      approval_policy: never
      inherit_process_env: true

  - id: anthropic:claude-agent-sdk
    label: Claude Code adapter
    config:
      model: claude-3-5-sonnet-latest
      working_dir: agent-workspace
      approval_policy: never
      sandbox_mode: read-only
      collect_trace: true
      trace_output: artifacts/claude-code-trace.jsonl
      env:
        ANTHROPIC_API_KEY: '{{env.ANTHROPIC_API_KEY}}'

uv run promptbeat eval \
  --config artifacts/coding-agent/generate/generated_redteam.yaml \
  --provider-file providers.comparison.yaml \
  --output-dir artifacts/comparison/eval
The generated probes are the same for every provider in the comparison. Only the eval step changes, which means differences in the report reflect real differences in agent behavior — not differences in the probes.
Keep provider files in version control so your target configs are auditable and reproducible. Add any .env files that contain real credentials to .gitignore. If you use a secrets manager, reference its export format in your shell setup script rather than in the provider YAML.