Files
openclaw/docs/tools/multi-agent-sandbox-tools.md
2026-04-29 11:54:28 +01:00

399 lines
12 KiB
Markdown

---
summary: "Per-agent sandbox + tool restrictions, precedence, and examples"
title: "Multi-agent sandbox and tools"
sidebarTitle: "Multi-agent sandbox and tools"
read_when: "You want per-agent sandboxing or per-agent tool allow/deny policies in a multi-agent gateway."
status: active
---
Each agent in a multi-agent setup can override the global sandbox and tool policy. This page covers per-agent configuration, precedence rules, and examples.
<CardGroup cols={3}>
<Card title="Sandboxing" href="/gateway/sandboxing">
Backends and modes — full sandbox reference.
</Card>
<Card title="Sandbox vs tool policy vs elevated" href="/gateway/sandbox-vs-tool-policy-vs-elevated">
Debug "why is this blocked?"
</Card>
<Card title="Elevated mode" href="/tools/elevated">
Elevated exec for trusted senders.
</Card>
</CardGroup>
<Warning>
Auth is scoped by agent: each agent has its own `agentDir` auth store at `~/.openclaw/agents/<agentId>/agent/auth-profiles.json`. Never reuse `agentDir` across agents. Agents can read through to the default/main agent's auth profiles when they do not have a local profile, but OAuth refresh tokens are not cloned into secondary agent stores. If you copy credentials manually, copy only portable static `api_key` or `token` profiles.
</Warning>
---
## Configuration examples
<AccordionGroup>
<Accordion title="Example 1: Personal + restricted family agent">
```json
{
"agents": {
"list": [
{
"id": "main",
"default": true,
"name": "Personal Assistant",
"workspace": "~/.openclaw/workspace",
"sandbox": { "mode": "off" }
},
{
"id": "family",
"name": "Family Bot",
"workspace": "~/.openclaw/workspace-family",
"sandbox": {
"mode": "all",
"scope": "agent"
},
"tools": {
"allow": ["read"],
"deny": ["exec", "write", "edit", "apply_patch", "process", "browser"]
}
}
]
},
"bindings": [
{
"agentId": "family",
"match": {
"provider": "whatsapp",
"accountId": "*",
"peer": {
"kind": "group",
"id": "120363424282127706@g.us"
}
}
}
]
}
```
**Result:**
- `main` agent: runs on host, full tool access.
- `family` agent: runs in Docker (one container per agent), only `read` tool.
</Accordion>
<Accordion title="Example 2: Work agent with shared sandbox">
```json
{
"agents": {
"list": [
{
"id": "personal",
"workspace": "~/.openclaw/workspace-personal",
"sandbox": { "mode": "off" }
},
{
"id": "work",
"workspace": "~/.openclaw/workspace-work",
"sandbox": {
"mode": "all",
"scope": "shared",
"workspaceRoot": "/tmp/work-sandboxes"
},
"tools": {
"allow": ["read", "write", "apply_patch", "exec"],
"deny": ["browser", "gateway", "discord"]
}
}
]
}
}
```
</Accordion>
<Accordion title="Example 2b: Global coding profile + messaging-only agent">
```json
{
"tools": { "profile": "coding" },
"agents": {
"list": [
{
"id": "support",
"tools": { "profile": "messaging", "allow": ["slack"] }
}
]
}
}
```
**Result:**
- default agents get coding tools.
- `support` agent is messaging-only (+ Slack tool).
</Accordion>
<Accordion title="Example 3: Different sandbox modes per agent">
```json
{
"agents": {
"defaults": {
"sandbox": {
"mode": "non-main",
"scope": "session"
}
},
"list": [
{
"id": "main",
"workspace": "~/.openclaw/workspace",
"sandbox": {
"mode": "off"
}
},
{
"id": "public",
"workspace": "~/.openclaw/workspace-public",
"sandbox": {
"mode": "all",
"scope": "agent"
},
"tools": {
"allow": ["read"],
"deny": ["exec", "write", "edit", "apply_patch"]
}
}
]
}
}
```
</Accordion>
</AccordionGroup>
---
## Configuration precedence
When both global (`agents.defaults.*`) and agent-specific (`agents.list[].*`) configs exist:
### Sandbox config
Agent-specific settings override global:
```
agents.list[].sandbox.mode > agents.defaults.sandbox.mode
agents.list[].sandbox.scope > agents.defaults.sandbox.scope
agents.list[].sandbox.workspaceRoot > agents.defaults.sandbox.workspaceRoot
agents.list[].sandbox.workspaceAccess > agents.defaults.sandbox.workspaceAccess
agents.list[].sandbox.docker.* > agents.defaults.sandbox.docker.*
agents.list[].sandbox.browser.* > agents.defaults.sandbox.browser.*
agents.list[].sandbox.prune.* > agents.defaults.sandbox.prune.*
```
<Note>
`agents.list[].sandbox.{docker,browser,prune}.*` overrides `agents.defaults.sandbox.{docker,browser,prune}.*` for that agent (ignored when sandbox scope resolves to `"shared"`).
</Note>
### Tool restrictions
The filtering order is:
<Steps>
<Step title="Tool profile">
`tools.profile` or `agents.list[].tools.profile`.
</Step>
<Step title="Provider tool profile">
`tools.byProvider[provider].profile` or `agents.list[].tools.byProvider[provider].profile`.
</Step>
<Step title="Global tool policy">
`tools.allow` / `tools.deny`.
</Step>
<Step title="Provider tool policy">
`tools.byProvider[provider].allow/deny`.
</Step>
<Step title="Agent-specific tool policy">
`agents.list[].tools.allow/deny`.
</Step>
<Step title="Agent provider policy">
`agents.list[].tools.byProvider[provider].allow/deny`.
</Step>
<Step title="Sandbox tool policy">
`tools.sandbox.tools` or `agents.list[].tools.sandbox.tools`.
</Step>
<Step title="Subagent tool policy">
`tools.subagents.tools`, if applicable.
</Step>
</Steps>
<AccordionGroup>
<Accordion title="Precedence rules">
- Each level can further restrict tools, but cannot grant back denied tools from earlier levels.
- If `agents.list[].tools.sandbox.tools` is set, it replaces `tools.sandbox.tools` for that agent.
- If `agents.list[].tools.profile` is set, it overrides `tools.profile` for that agent.
- Provider tool keys accept either `provider` (e.g. `google-antigravity`) or `provider/model` (e.g. `openai/gpt-5.4`).
</Accordion>
<Accordion title="Empty allowlist behavior">
If any explicit allowlist in that chain leaves the run with no callable tools, OpenClaw stops before submitting the prompt to the model. This is intentional: an agent configured with a missing tool such as `agents.list[].tools.allow: ["query_db"]` should fail loudly until the plugin that registers `query_db` is enabled, not continue as a text-only agent.
</Accordion>
</AccordionGroup>
Tool policies support `group:*` shorthands that expand to multiple tools. See [Tool groups](/gateway/sandbox-vs-tool-policy-vs-elevated#tool-groups-shorthands) for the full list.
Per-agent elevated overrides (`agents.list[].tools.elevated`) can further restrict elevated exec for specific agents. See [Elevated mode](/tools/elevated) for details.
---
## Migration from single agent
<Tabs>
<Tab title="Before (single agent)">
```json
{
"agents": {
"defaults": {
"workspace": "~/.openclaw/workspace",
"sandbox": {
"mode": "non-main"
}
}
},
"tools": {
"sandbox": {
"tools": {
"allow": ["read", "write", "apply_patch", "exec"],
"deny": []
}
}
}
}
```
</Tab>
<Tab title="After (multi-agent)">
```json
{
"agents": {
"list": [
{
"id": "main",
"default": true,
"workspace": "~/.openclaw/workspace",
"sandbox": { "mode": "off" }
}
]
}
}
```
</Tab>
</Tabs>
<Note>
Legacy `agent.*` configs are migrated by `openclaw doctor`; prefer `agents.defaults` + `agents.list` going forward.
</Note>
---
## Tool restriction examples
<Tabs>
<Tab title="Read-only agent">
```json
{
"tools": {
"allow": ["read"],
"deny": ["exec", "write", "edit", "apply_patch", "process"]
}
}
```
</Tab>
<Tab title="Safe execution (no file modifications)">
```json
{
"tools": {
"allow": ["read", "exec", "process"],
"deny": ["write", "edit", "apply_patch", "browser", "gateway"]
}
}
```
</Tab>
<Tab title="Communication-only">
```json
{
"tools": {
"sessions": { "visibility": "tree" },
"allow": ["sessions_list", "sessions_send", "sessions_history", "session_status"],
"deny": ["exec", "write", "edit", "apply_patch", "read", "browser"]
}
}
```
`sessions_history` in this profile still returns a bounded, sanitized recall view rather than a raw transcript dump. Assistant recall strips thinking tags, `<relevant-memories>` scaffolding, plain-text tool-call XML payloads (including `<tool_call>...</tool_call>`, `<function_call>...</function_call>`, `<tool_calls>...</tool_calls>`, `<function_calls>...</function_calls>`, and truncated tool-call blocks), downgraded tool-call scaffolding, leaked ASCII/full-width model control tokens, and malformed MiniMax tool-call XML before redaction/truncation.
</Tab>
</Tabs>
---
## Common pitfall: "non-main"
<Warning>
`agents.defaults.sandbox.mode: "non-main"` is based on `session.mainKey` (default `"main"`), not the agent id. Group/channel sessions always get their own keys, so they are treated as non-main and will be sandboxed. If you want an agent to never sandbox, set `agents.list[].sandbox.mode: "off"`.
</Warning>
---
## Testing
After configuring multi-agent sandbox and tools:
<Steps>
<Step title="Check agent resolution">
```bash
openclaw agents list --bindings
```
</Step>
<Step title="Verify sandbox containers">
```bash
docker ps --filter "name=openclaw-sbx-"
```
</Step>
<Step title="Test tool restrictions">
- Send a message requiring restricted tools.
- Verify the agent cannot use denied tools.
</Step>
<Step title="Monitor logs">
```bash
tail -f "${OPENCLAW_STATE_DIR:-$HOME/.openclaw}/logs/gateway.log" | grep -E "routing|sandbox|tools"
```
</Step>
</Steps>
---
## Troubleshooting
<AccordionGroup>
<Accordion title="Agent not sandboxed despite `mode: 'all'`">
- Check if there's a global `agents.defaults.sandbox.mode` that overrides it.
- Agent-specific config takes precedence, so set `agents.list[].sandbox.mode: "all"`.
</Accordion>
<Accordion title="Tools still available despite deny list">
- Check tool filtering order: global → agent → sandbox → subagent.
- Each level can only further restrict, not grant back.
- Verify with logs: `[tools] filtering tools for agent:${agentId}`.
</Accordion>
<Accordion title="Container not isolated per agent">
- Set `scope: "agent"` in agent-specific sandbox config.
- Default is `"session"` which creates one container per session.
</Accordion>
</AccordionGroup>
---
## Related
- [Elevated mode](/tools/elevated)
- [Multi-agent routing](/concepts/multi-agent)
- [Sandbox configuration](/gateway/config-agents#agentsdefaultssandbox)
- [Sandbox vs tool policy vs elevated](/gateway/sandbox-vs-tool-policy-vs-elevated) — debugging "why is this blocked?"
- [Sandboxing](/gateway/sandboxing) — full sandbox reference (modes, scopes, backends, images)
- [Session management](/concepts/session)