Strands Engine

Libre WebUI embeds an agent engine built on the open-source Strands Agents harness (@strands-agents/harness 0.1.1 on @strands-agents/sdk 1.19.0). It runs inside the Libre WebUI backend process, so there is no separate daemon to install. The SDK is loaded the first time the engine is used; a server with Strands switched off never loads it.

The engine drives the models Libre WebUI already serves:

• Ollama models, when Ollama is enabled
• models from active chat or completion provider plugins

Every model call goes through the same Ollama and plugin services Chat uses. There is no separate provider configuration and no separate API key. Provider credentials and the Ollama switch apply unchanged.

You can use the engine in three places:

• the Strands page, for persistent agent sessions
• Chat, through the Strands entries in the model selector
• Work, as the Strands option of the composer's Engine control

Access

An administrator chooses who can use the engine under Settings → User Management → Access & policies → Strands engine:

Mode	Label in the setting	Who can use Strands
disabled	Off	Nobody, administrators included (default)
admins	Administrators	Active administrators
all-users	All users	Every active account

A change takes effect on the next request, without a restart.

To pin the mode at the deployment level, set LIBRE_STRANDS_ACCESS:

LIBRE_STRANDS_ACCESS=admins   # or disabled, all-users

A pinned mode locks the control in User Management. Any value other than disabled, admins, or all-users locks the engine off instead of falling back to the saved setting.

The server enforces the mode on every REST request, on WebSocket chat, and on every model call of a Work run. If the mode cannot be read, access is denied. Accounts without access do not see the Strands page, the Strands chat entries, or the Strands Work engine, and the API answers 403.

Strands page

Open Strands from the sidebar (/strands).

• Sessions persist across restarts, up to 200 per account. The first message names a new session.
• Pick a model for each session. Default model uses your default chat model.
• Replies stream as they are generated, with the model's reasoning and a card for each tool call.
• Stop cancels the running turn. A session runs one turn at a time.
• Deleting a session removes its history and its workspace.

Chat

Pick Strands in the model selector. It appears in the Agents group next to the Agent CLI models:

• Strands (strands) uses the engine default, which is your default chat model.
• Strands · model (provider) (strands:<route>) pins one model and its provider for the conversation.

Strands in Chat depends only on Strands access. It does not need the Agent CLI models toggle, and it appears while that toggle is off. In Chat the agent has the same file tools as on the Strands page, in one private workspace per account, and its replies arrive as text and reasoning. See Installed coding agents for the Agents group and for using a Strands entry as the task model for titles and thinking summaries.

Work

In the Work composer, choose Libre WebUI or Strands in the Engine control and pick the model separately. With Strands, a Strands agent plans each step, and Work executes the tools it asks for in the Work sandbox under Work's normal approval policy. Work stays authoritative for the transcript, approvals, tool execution, and run history; the Strands agent runs no tools of its own in Work.

The underlying provider model must support tool calling. Work checks this before a run starts and rejects a model that does not advertise tool support.

Work runs saved by earlier releases with a dsh: model prefix open as Strands runs. See Work: Strands engine for how the engine choice interacts with models carried over from Chat.

Security model

On the Strands page and in Chat, the agent has a deliberately small tool set:

• read, write, and edit, for files in its workspace
• the harness todos plugin, for planning multi-step work

It has no shell, no web fetch, no memory, and no skills, and no AGENTS.md file is injected into its instructions.

Each session has a private workspace directory, which the agent sees as /workspace. Every path is resolved after following symlinks and must stay inside that directory, so a link in the workspace cannot reach the rest of the disk. The sandbox refuses to execute commands.

Item	Limit
Sessions per account	200
Prompt length	32,000 characters
Agent turns per message	24
File read or write	2 MB per call
Session title	120 characters

A turn that reaches the turn limit stops and says so. Each turn can make several model calls to your providers, so grant access with the same care as any other model access that spends provider budget.

Storage

Strands data lives under <DATA_DIR>/strands/, not in the application database. Each account has its own directory, <DATA_DIR>/strands/users/<account-hash>/, containing:

• registry.json: the account's session list
• sessions/: the Strands session snapshots the agent uses as context
• transcripts/: the transcript shown on the Strands page for each session
• workspaces/: the private workspace of each session
• chat-workspace/: the workspace used by Chat turns

Include <DATA_DIR>/strands/ in backups if the sessions matter to you.

Usage

Provider Usage (/usage) attributes Strands calls to the Strands agent in its Agents section. See Agent usage.

API

Every route requires an authenticated account with Strands access, except /api/strands/access, which requires an administrator.

Method	Path	Purpose
GET	/api/strands/access	Read the access mode and whether LIBRE_STRANDS_ACCESS pins it (admin)
PUT	/api/strands/access	Set mode to disabled, admins, or all-users (admin; 409 if pinned)
GET	/api/strands/health	Engine availability and the harness and SDK versions
GET	/api/strands/models	Models the engine can drive for this account
GET	/api/strands/sessions	List sessions
POST	/api/strands/sessions	Create a session (optional title and model)
GET	/api/strands/sessions/:sessionId	Read a session, its transcript, and whether a turn is running
PATCH	/api/strands/sessions/:sessionId	Change a session's title or model
DELETE	/api/strands/sessions/:sessionId	Delete a session with its transcript and workspace
POST	/api/strands/sessions/:sessionId/messages	Send { "text": "..." } and stream the turn as NDJSON
POST	/api/strands/sessions/:sessionId/cancel	Stop the running turn

The message stream writes one JSON event per line: turn-start, text, reasoning, tool-start, tool-result, done (with the stop reason and token usage when the provider reports it), and error. Closing the connection cancels the turn. Sending a message while a turn is running returns 409.

Configuration

Variable	Default	Purpose
LIBRE_STRANDS_ACCESS	unset (admin setting, admins)	Pin disabled, admins, or all-users; any other value locks the engine off

Troubleshooting

See Strands Engine Problems.