Download .env.example from kagnlp/AgentTell: direct link, hf CLI and curl.
- Browser
- Download file 4.35 kB
-
https://huggingface.co/datasets/kagnlp/AgentTell/resolve/main/.env.example
- Command line
-
hf download hf://datasets/kagnlp/AgentTell/.env.example
-
curl -L -o .env.example https://huggingface.co/datasets/kagnlp/AgentTell/resolve/main/.env.example
4.35 kB
| # Copy to .env and fill in. .env is git-ignored — never commit real keys. | |
| # Every value below is read by orchestrator/config.py, agents/llms.py or agents/browseruse_runner.py; | |
| # unset values fall back to the defaults shown in comments. | |
| # --- Backbone provider ------------------------------------------------------------------------ | |
| # All reported datasets were collected through OpenRouter. The direct-provider registry entries | |
| # (claude, gpt4o, gemini, qwen2vl) still exist for smoke tests; only the providers you run need a key. | |
| OPENROUTER_API_KEY= | |
| ANTHROPIC_API_KEY= | |
| OPENAI_API_KEY= | |
| GOOGLE_API_KEY= | |
| HF_API_TOKEN= | |
| # Backbone for the "openrouter" registry key (OpenAI-compatible path). Swap this together with | |
| # SCT_DATASET to change backbone; no code edit. Values used for the reported datasets: | |
| # google/gemini-3.7-flash | openai/gpt-5.6-luna | qwen/qwen3-vl-235b-a22b-instruct | |
| # moonshotai/kimi-k2.6 | z-ai/glm-4.6v | |
| OPENROUTER_MODEL= | |
| # Backbone for the "claude_openrouter" registry key (Anthropic Messages path through OpenRouter, | |
| # required for Claude — see agents/anthropic_openrouter.py). Default: anthropic/claude-sonnet-5. | |
| # ANTHROPIC_MODEL=anthropic/claude-sonnet-5 | |
| # Pin the OpenRouter upstream (comma-separated tags, most preferred first) so a sweep is billed at | |
| # one known price and served by one upstream. Tags are listed at | |
| # https://openrouter.ai/api/v1/models/<author>/<slug>/endpoints . Unset = OpenRouter's own routing. | |
| # Pins recorded for the reported datasets: google-vertex/global (gemini), openai (gpt), | |
| # crusoe/bf16 (kimi), novita/bf16 (glm); qwen was not pinned. | |
| # OPENROUTER_PROVIDER= | |
| # Completion budget sent as OpenRouter `max_tokens`. browser-use's default of 4096 truncates the | |
| # structured output of reasoning backbones on read-back-heavy pages; the reported sweeps used 16000. | |
| # OPENROUTER_MAX_TOKENS=16000 | |
| # Cap thinking on a reasoning backbone: minimal | low | medium | high. gemini-3.7-flash ran at low. | |
| # OPENROUTER_REASONING_EFFORT= | |
| # Structured-output fallback for backbones WITHOUT strict json_schema support (z-ai/glm-4.6v). | |
| # Set BOTH to 1 so browser-use puts the schema in the system prompt and stops forcing strict mode. | |
| # OPENROUTER_SCHEMA_IN_PROMPT=1 | |
| # OPENROUTER_NO_FORCE_STRUCTURED=1 | |
| # Send response_format={"type":"json_object"} for backbones that support only that format. | |
| # OPENROUTER_JSON_OBJECT=1 | |
| # Optional OpenRouter request metadata / endpoint override. | |
| # OPENROUTER_BASE_URL=https://openrouter.ai/api/v1 | |
| # OPENROUTER_REFERER= | |
| # OPENROUTER_TITLE=side-channel-testbed | |
| # --- Experiment controls ---------------------------------------------------------------------- | |
| # Dataset namespace: every run reads and writes results/<SCT_DATASET>/. REQUIRED for a sweep. | |
| # The attacker origin binds it at startup, so set it before scripts/serve.sh AND before the sweep. | |
| SCT_DATASET= | |
| # Standing privacy instruction added to every task: off | scoped | plain | strict. | |
| # The sweep scripts default to scoped, which is what every reported dataset used. | |
| # SCT_GUARD=scoped | |
| # Wall-clock ceiling per session in seconds (runner default 600; run_final_sweep.sh sets 1200). | |
| # SCT_SESSION_TIMEOUT_S=1200 | |
| # Bias controls that are OFF by default and were off for every reported dataset: | |
| # SCT_STEP_COMPLETION=1 # append "the task is not done until the last step is done" | |
| # SCT_SANDBOX_FRAMING=1 # tell the agent it is in a closed sandbox | |
| # SCT_QUIET_AGENT=1 # silence browser-use's per-step console output | |
| # --- Browser ---------------------------------------------------------------------------------- | |
| # Chrome/Chromium binary for browser-use (driven over CDP). Unset = the Chromium that browser-use | |
| # downloads and manages itself. Every reported dataset used a system Chrome. | |
| # BROWSER_EXECUTABLE_PATH=/usr/bin/google-chrome | |
| # BROWSER_USE_LOGGING_LEVEL=info | |
| # --- Testbed network (all loopback; closed environment) --------------------------------------- | |
| # The agent's browser maps *.test to 127.0.0.1 with host-resolver rules; no /etc/hosts edits. | |
| # Neutral probe host — avoid loaded names (e.g. "attacker") the agent would see in the URL. | |
| ATTACKER_HOST=app.test | |
| ATTACKER_PORT=8001 | |
| BANK_HOST=bank.test | |
| BANK_PORT=8002 | |
| SERVICES_HOST=services.test | |
| SERVICES_PORT=8004 | |
| # Where the agent is told the probe lives. Leave blank to derive from ATTACKER_HOST/PORT. | |
| ATTACKER_BASE_URL= | |