Jarela
Jarela is a local-first desktop UI for orchestrating LangGraph agents on your own machine. It keeps agent state, memory, schedules, and model/tool configuration in SQLite under a local data directory, with no hosted backend required.
New to Jarela? Visit the documentation site for install, operations, architecture, and extension guides.
Highlights
- Local-first, single-process Next.js app
- Multi-provider LLM support: Anthropic, OpenAI, Google GenAI, Cohere, DeepSeek, GitHub Copilot
- Per-agent tool policy, identity, memory, and harness configuration
- Wide-screen agent sidebar with per-agent model, usage strategy, and voice controls
- Per-agent Home, Work, Developer, and Everything usage profiles for response style, chat diagnostics, and credential discovery
- SQLite-backed checkpoints, memory, schedules, proposals, and local integrations
- Durable, agent-targeted lifecycle signals for restarts, approvals, configuration changes, and background work
- PWA / desktop-friendly workflow with browser extension support
- MCP, integration manifests, and built-in tool extension points
- Mermaid diagram previews with expandable, copyable source and sanitized SVG downloads
- Guided first-run setup that helps you create a default assistant and personalize it in chat
Jarela in action
First launch
Follow the first-run guide for the full setup flow and what to expect in your first chat.
Dashboard costs (Jarela 3.0.0)
The dashboard shows separate input and output prices per million tokens for each model, and separate input/output spending in the summary, daily selection, and vendor/model/agent reports. Input spending includes prompt-cache read/write charges under the existing provider pricing assumptions. Costs are current-rate estimates, not invoices; recorded historical totals remain unchanged. When no matching rates are available, a recorded cost is shown as unattributed instead of inventing an input/output split. Legacy messages still use content-length token estimates.
Document indexing and recall privacy (Jarela 3.0.0)
In Documents, use Opt in to indexing to choose a local folder or a supported connected integration (Gmail, Outlook, iCloud, GitHub, Jira, or Confluence). Scope remote sources to a mailbox, repository, project, or space. Connecting credentials does not automatically index content. Removing a source deletes its indexed documents and chunks without changing the originals.
Gmail, Outlook, and iCloud mail sources offer Inbox, a chosen folder or
label, or whole mailbox (including spam and trash). Search is optional.
For chosen folders, Gmail accepts an exact label name or ID; Outlook accepts a
folder ID or well-known name such as sentitems; iCloud accepts a mailbox path.
Leave Max results blank for all messages, or set a positive limit. Outlook
text search has a Microsoft Graph ceiling of 1,000 results; leave search blank
to scan a complete folder/mailbox. New sources keep an archive: previously
indexed messages remain even if moved, deleted, or no longer matching the
query. Enabled sources refresh approximately every ten minutes while Jarela is
running and unlocked; this is polling, not instant delivery. Large initial
scans may take longer. Existing query-only Gmail/Outlook sources retain their
previous bounded mirror behavior; recreate them to use the new scopes and
archive defaults. iCloud indexing requires the updated iCloud package.
iCloud mail sending
The updated iCloud package also provides icloud_mail_send_email, classified
as execute like Gmail/Outlook sending. Enable it explicitly in the agent's
tool permissions and only request it when you want to send immediately; use
drafts for review. It reuses the saved app-specific password. If your Apple ID
differs from your mail address, set Credentials → iCloud → iCloud Mail
address (or ICLOUD_MAIL_ADDRESS with environment credentials). SMTP uses
the full mail address, with mandatory STARTTLS on port 587. A successful send
means SMTP acceptance, not confirmed delivery. Do not automatically retry
uncertain or partially accepted sends. See the package sending guide.
Indexing with an embedding provider other than Jarela Local displays a warning and requires confirmation in the source setup flow. Content may be sent to that provider, including material later excluded from recall. Switching away from Jarela Local also requires confirmation. Locally embedded excerpts can still be sent to a cloud chat model when retrieved for an answer.
Global recall exclusions apply to all indexed sources, both existing and new:
- Save topic descriptions for automatic, best-effort semantic filtering of matching chunks. The default similarity cutoff is 0.8; lower values exclude more. Topic descriptions use the selected embedding provider.
- Search a topic and choose Exclude document to block all chunks of a selected document. This exact exclusion uses source ID and path, so it survives reindexing within that source. Recreating a source requires selecting its document exclusions again.
- Assistant document searches respect exclusions and cannot enable an override. The user-facing search preview has a separate Include excluded content checkbox, off by default, which does not grant the assistant access.
- If topic vectors or chunk embeddings are unavailable, invalid, or have incompatible dimensions or model signatures, filtered search fails explicitly rather than returning unchecked chunks. Reindex existing sources before using topic rules for the first time after upgrading: older vectors have no model provenance. Also reindex sources after changing embedding models.
Semantic filtering is not a privacy guarantee: mixed-topic chunks and indirect references can be missed. Search previews are ranked and limited, not exhaustive audits. Exclusions do not delete stored content, prevent external embedding, remove excerpts already in chats, or restrict separate live email/file tools.
Model updates (Jarela 3.0.0)
The built-in pricing-maintenance agent tool also checks official model documentation for version and capability changes. It refreshes catalog/badge metadata without switching saved models or enabling new APIs; confirmed adoption gaps are escalated through deduplicated GitHub issues when access and approval are available. See model-update operations.
Agent usage profiles
Jarela 3.0.0: On wide screens, the scrollable left sidebar shows agent names, unread counts, usage profiles, selected-tool counts, and model routing. The selected-agent card above the right menu provides quick model, usage-strategy, and voice controls. These settings save immediately; model and strategy changes apply to new runs. Narrow screens use the mobile header and menu drawer instead of docked sidebars. The phone image above uses an illustrative device frame around a mobile-width capture, not a verified installed iOS PWA screenshot. See desktop agent controls.
The top-left agent switcher keeps the current agent first, followed by agent icons that fade with distance. Hover or keyboard focus restores their opacity; selecting one brings it forward and opens its chat with an animated handoff. The strip scrolls horizontally on small screens and respects reduced motion. Background icons retain the existing default/latest-message ordering. Unread notifications use shared per-agent tracking and appear as bright dots, independent of icon fading; opening the agent's chat clears its unread state.
The menu is destination-based rather than a second agent picker. Workspace
contains Agents, plus Knowledge, Automations, and Usage in Full mode.
Profile and Settings live under Preferences; Capabilities is under the
collapsible Advanced section in Full mode. Models, credentials, and logs remain
in Settings. Full-width, touch-friendly rows use complete labels on small screens
and close the menu after navigation. Switch agents or return to chat using the
header strip; selecting an agent there also closes the menu.
Public branding values, including unset values, are inlined at build time so
default logos and names do not require a browser-side Node process.env shim.
Choose a Persona under an agent's Identity settings:
Selecting a persona applies editable style and citation defaults. Tune warmth, expressiveness, response detail, and conversational-cue adaptation in the same Persona section; citation coverage stays independent in Advanced settings. Zero cue adaptation keeps style steady. Disabling style customization retains saved sliders and the profile's organization. MBTI has been removed from the editor and prompt behavior; existing tuned slider values are preserved, with no automatic reset on upgrade or save. Home and Work keep Technical details inside chat's single expandable Filters & focus control, alongside channel and focus controls. Failures, approval requests, source links, and factual-accuracy requirements remain visible or applicable in every profile. Diagnostics are still collected locally; the profile changes presentation, not tool permissions or model-context filters. Actual model adherence to style and grounding instructions depends on the provider.
System prompts separate instruction composition, agent identity/instructions, navigation, a single resolved Persona, and grounding/citations into named chapters. The Persona combines the selected profile's organization and direct style targets; explicit task depth takes priority. Conversational cues are tentative heuristics, not emotion or personality diagnoses. Citation coverage follows the independently editable citation setting. Assumptions never count as verified evidence, and resource-saving modes cannot skip required validation. Plain-text delivery channels use source URLs instead of Markdown citation wrappers.
The Credentials catalog follows the active agent's profile. Configured integrations and saved credentials stay accessible; Show all integrations bypasses discovery filtering. With no active agent, the catalog shows everything. Global Essential/Full mode still controls app configuration panels independently.
Web search and fetch backends (Jarela 3.0.0)
web_search and web_fetch are Jarela tools, separate from model-native web-search
capabilities shown in model badges. Firecrawl Search and Jarela's guarded direct
page fetch remain the defaults. To opt in to DeepSeek's separate Search and Fetch
APIs, configure DeepSeek Search & Fetch under Settings → Credentials. This
dedicated encrypted credential is required; configuring a DeepSeek chat-model key
alone never sends web queries or URLs to DeepSeek. Removing the dedicated credential
restores the existing backends.
DeepSeek Fetch returns Markdown without image extraction. Raw HTML requests, hosts in Allowed sites (including cookie-authenticated pages), and private-fetch overrides retain the existing direct-fetch path. Jarela checks URLs against its SSRF policy before using DeepSeek Fetch. As documented by DeepSeek on 2026-10-25, successful Search and Fetch requests are each priced at $0.75 per 1,000 requests; empty searches and failed requests are not billed. See the DeepSeek Search & Fetch guide for current limits, pricing, and API behavior.
On upgrade, the old user-profile Persona preset is copied once to existing agents, without rewriting their tuned personality or citation settings. New agents default to Everything. See ADR-0105.
Quick start
npm install
npm run dev
Then open http://127.0.0.1:4312.
Documentation
Full docs are organized under the docs site and source docs folder:
- Live documentation site
- Docs home
- Install guide
- First-run guide
- Development guide
- Architecture
- Operations guide
- Extending Jarela
- API reference
- Contributing
Maintainers can optionally run Copilot to propose a version-to-version documentation PR. Releases do not require this review or a Copilot token; ordinary CI and nightly releases never start paid AI sessions. See the on-demand release documentation workflow. This is development/release infrastructure, not a cloud dependency of the app.
Modular design
Jarela is a single Next.js process organized around domain modules and explicit runtime boundaries:
- The UI calls typed API routes; route handlers guard and validate requests before delegating to
lib/. lib/agents/coordinates turn execution through provider and tool registries, with stores owning persistence.- Providers, built-in tools, MCP servers, bridges, and integrations connect through adapters and registries instead of reaching into each other's implementations.
- Background work reuses the existing scheduler, agent queues, SQLite outbox, and leased signal delivery; there is no additional broker or agent process.
See the C4 architecture and module map for runtime relationships, and the development guide for module-specific extension rules.
Extension surfaces
Jarela supports explicit extension points for:
- built-in LLM providers
- external provider plugins
- built-in tools
- MCP servers
- custom harnesses
- integration manifests
- app branding / runtime config
Use the agent tool describe_extension_surfaces and the docs guide for the correct entry point before editing frameworks or config.
Repository structure
Project commands
npm run dev
npm run build
npm run lint
npm test
npm run docs:dev
npm run docs:build
The docs commands use Zensical; install it with
pip install -r requirements-docs.txt (Python 3.11 or newer).
License
Apache-2.0