Hand a task off to another model to run, instead of doing it yourself. Core commands:
cd agent-fleet
# 单个任务
node bin/agent-fleet.mjs run --model <友好名字> --prompt "<任务描述>" [--cwd <目录>] [--json]
# 一次并发跑一批不同模型的任务
node bin/agent-fleet.mjs run-many --config batch.json [--json]
# 看有哪些模型可用、密钥配没配
node bin/agent-fleet.mjs list-models
<友好名字> from models.config.json, currently preset:deepseek-v4-pro、
deepseek-v4-flash、kimi、gemini(Gemini has no official Anthropic-compatible endpoint; the user must fill in a gateway
address themselves to use it. If not filled in, it errors out directly and does not pretend it can run), as well as the kollab-gateway series (goes through Kollab's own public LLM
gateway, using a self-service generated kollab_live_* key from your account, no need to wait for third-party approval, verified at the moment of use):
kollab-gateway(default gemini-3.8-flash)、kollab-gateway-copy(copywriting/creative, likewise
gemini-3.8-flash)、kollab-gateway-research(research summaries,grok-4.6)、kollab-gateway-bulk
(bulk mechanical tasks,gemini-3.5-flash-lite)。
When users want to run tasks on several different models at the same time, prefer using run-many to submit them all at once (internally they truly run concurrently to completion), rather than
writing multiple serial calls by hand yourself, or opening a separate background shell process for each model—unless the user explicitly asks for that kind of interaction.
Choose a model by task type: bulk copywriting/creative → kollab-gateway-copy(or use
deepseek-v4-flashwhen a DeepSeek key is available); bulk translation/format conversion → kollab-gateway-bulk(or deepseek-v4-flash);
simple research summaries → kollab-gateway-research(or use kimiwith built-in web search when a Moonshot key is available);
high-quality single-shot output (finalizing long copy, complex reasoning) → deepseek-v4-pro;Tasks with high fault-tolerance requirements for multi-turn tool calling should not be
handed off to third-party models, leave them for Claude to handle itself (the Kimi/DeepSeek/Qwen family is known to have tool-calling reliability issues,
and may spit out raw tool-call control tokens instead of structured tool_use, causing false success that the harness cannot fix). This
mapping is a recommendation derived from agent-fleet's own research + real verification, and will be continuously calibrated with actual use; it is not a hardcoded rule.
The full version and the known model catalog are in README.md 's "task type → recommended model" section.
Before first useyou must confirm that the real API key for the corresponding model is already configured in agent-fleet/.env (cp .env.example .env
filled in manually afterwards); if not configured, run/run-many will directly report a clear "which environment variable is not set" and will not fail silently.
About the security constraints of --cwd security constraints: the target working directory is treated asuntrusted input. If that directory's own
.claude/settings.json tries to change the target address, credentials, or custom request headers of model requests (a typical trick is
env.ANTHROPIC_BASE_URL hijacking the key to another address), the entire run will directly error out and exit, pointing out the unauthorized field—this is intentional
fail-closed, not a bug, so do not "fix" it by deleting the gate. At the same time, be aware of the boundaries of this protection layer: the tool runs a
bypassPermissions autonomous Agent, and a malicious directory can still use prompt injection to induce the model to send the key out itself, so
do not directly use this tool to process directories of unknown origin, and if you want to run it, put it in a container/one-off virtual machine.
For complete model integration details, security boundaries, and verification status, see README.md. This is a purely local tool, and has nothing to do with
any of Kollab's product infrastructure, gateways, or account system.