Retries & circuit breakers
Exponential backoff with wall-time caps, 429 and 5xx budgets, provider-level circuit breakers that route around degraded endpoints.
Feature list
500+ herramientas. Un harness.
Edición de archivos, terminal, git, navegador, memoria, documentos, integraciones y flujos de trabajo multiagente: con puertas de aprobación y ejecución local.
The harness
A loop engineered to keep running
Liminal is a structured ReAct loop with production plumbing built to survive the failure modes that kill most agents in the wild.
Context compression
Hot rounds verbatim, warm rounds keep provenance, stale tool results elided to artifact pointers in .agent_artifacts/.
Drift scoring
Per-turn execution contracts track whether the agent is still on its plan. Detected drift triggers reflection, not silent wandering.
Approval gates
Destructive tools require explicit approval in the UI. Optional safety judge skips approval for provably safe calls.
Resumable streaming writes
Large write_file / edit_file payloads stream into staging. Length-truncated writes resume from a manifest.
Compensation ledger
Records partial side effects so a failed multi-step plan can be reversed.
Tools
500+ tools, organized into families
Lazy-loaded by default. The model sees a baseline set; specialized families activate on demand.
Files & code
read_file, edit_file, write_file, multi_file_apply, ast_grep, symbol_index, find_references, rename_symbol.
Shell & process
run_shell (approval-gated), run_background, run_tests, run_lint.
Git
git_status through git_worktree for linked branches.
Web
web_search + web_fetch with readability extraction and bot-wall retries.
Browser automation
Playwright-backed navigate, snapshot, act, extract, captcha_solve.
Memory & vault
remember, recall_relevant, memory_graph, vault_read/write/search.
Reasoning
think, reason, plan, hypothesize, breakdown (observable, not hidden).
Documents
doc_plan through doc_render for PPTX, DOCX, PDF.
Voice
Dictate tasks and hear responses with built-in transcription and text-to-speech.
Workflows (harness)
plan_workflow / run_workflow (local EE only, not a cloud API).
Meta
spawn_agent, model-defined dynamic tools, OpenAPI and MCP connectors.
Integrations
One-click connectors, no app registration
Connect your tools from the Liminal Desktop app through Vireon's hosted OAuth bridge. No per-user OAuth app to register and no client secrets in your .env.
Google Workspace
Gmail, Docs, Sheets, Slides, Calendar (REST + MCP).
Microsoft 365
Outlook, Teams, Planner, To Do, OneNote, OneDrive, Excel, SharePoint.
GitHub
Repos, issues, and pulls via the GitHub connector.
Slack & Linear
Messages and issue tracking through hosted OAuth.
Notion & Xero
Docs/databases and accounting through hosted OAuth.
Hosted token refresh
Long-lived connections refresh server-side; secrets never touch your machine.
Memory & knowledge
Persistence that compounds across sessions
Memory in Liminal isn't a vector dump. It's typed, ranked, contradicted, curated, and consolidated.
Typed notes
Notes carry a type and route to ranking strategies tuned for each.
Hybrid recall
BM25 + embedding recall with recency and type bias.
Cross-chat federation
Notes are scoped per-chat, per-workspace, or global, so the right context follows you between sessions without bleeding everywhere.
Recipe library
Successful multi-tool turns recorded by intent and tool-phase shape.
Memory curator (harness)
LLM prune-and-merge with guardrails. Runs in the local agent, not on the cloud control plane.
Obsidian vault
Read/write against your vault. See the Obsidian use case.
Auto-dream
Background consolidation during idle time (off by default).
Persona system
One prompt generates a whole assistant
First-run persona bootstrap builds voice, behaviors, and a validated UI theme for the web shell.
Soul slices
Voice, motivations, mannerisms, taboos, and a living scratch.
UI themes
PersonaUiTheme drives Terminal / Hud / Studio / Minimal shells.
Heartbeat
Optional idle-time autonomous ticks.
Streaming generation
Artifacts stream live during bootstrap.
Reasoning & routing
Cheap when it can be, deep when it has to be
Per-turn intent classification picks the right model and reasoning budget, and routing chooses the cheapest healthy provider.
Two-tier models
Main model runs the ReAct turn. Fast model runs classification and sidecars.
Managed inference
Optional on Pro: route chat through Vireon's metered OpenAI-compatible proxy. No API key required, with included monthly credits.
Dynamic provider routing
Price-sorted across OpenRouter resellers, sticky sessions for KV-cache affinity, and automatic rotation around rate-limited providers.
Reasoning budget
External surface: think() + reason() tools (observable).
Prompt caching
~13× cheaper cached prefix on supported providers.
Effort learning
Outcome scoring reuses best effort per intent.
Observability
See every step the loop takes
Liminal Desktop renders every tool call inline in a polished chat UI. CLI and browser UIs expose the same harness for developers. JSONL traces capture everything.
Live tool cards
Each tool call with args, status, and result.
Streaming text
Token-level streaming; SSE reconnect on web.
Session JSONL
.agent_sessions/<taskId>.jsonl (replayable).
Self-telemetry
Harness failure log, rule stats, recipe stats.
Settings UI
~190 harness knobs in desktop Settings (web Settings for CLI path) → .agent_runtime_prefs.json.
Pro & cloud
Optional cloud, when you want it
Community Edition is fully local and needs no account. Sign in for Pro to add metered managed inference and sync across machines. Your code still never touches our servers.
Managed inference
Run without your own provider key. Metered, OpenAI-compatible proxy with included monthly credits and optional top-ups.
Cloud memory sync
Typed notes and your vault sync across machines, so the agent's memory is the same everywhere you work.
Session history
Searchable, cloud-stored session history on the control plane.
Team (cloud)
Org admin & governance APIs
Team adds hosted shared memory and a web console at /account/organization: members, invites, audit export, fleet JSON, and policy JSON. The harness ingests audit events and applies policy on login via Enterprise Edition.
Shared memory bus
Hosted bus per workspace/org for squad coordination.
Audit log
Harness ships turn summaries; admins query and export CSV in the account UI.
Org RBAC
Owner, admin, member, viewer. Role changes, invites, and seat limits tied to Stripe.
Fleet & policy
Revisioned JSON stored on the control plane. EE enforces policy on each machine.
Enterprise (assisted)
SSO-ready IdP setup and self-host evaluation with sales. Productized SAML/SCIM and packaged CP deploy are on the roadmap.
Evaluate
Compare & setup guides
Honest comparisons vs Cursor, Copilot, and Claude Code, plus install and OpenRouter guides.
Features FAQ
Liminal AI feature questions
How many tools does Liminal AI have?
Liminal ships with 500+ built-in tools organized into families: files and code intelligence, shell and processes, git, web search and fetch, Playwright browser automation, memory and Obsidian vault, document generation, integrations, and reasoning. Tools are lazy-loaded so the model only sees the families a task needs.
Can Liminal AI edit files and run terminal commands?
Yes. Liminal reads and edits files (whole-file or fuzzy-diff), runs shell commands and background processes, executes tests and linters, and commits to git. Destructive commands are approval-gated in the UI before they run.
Does Liminal integrate with git?
Yes. It has first-class git tools: status, diff, log, branch, commit, checkpoint, rollback, and linked worktrees.
Does Liminal remember context between sessions?
Yes. Typed, ranked memory with hybrid BM25 + vector recall, recipe library, and optional Obsidian vault sync.
Can Liminal automate a web browser?
Yes. Playwright-backed browser tools with optional CAPTCHA integration.
Does Liminal support voice?
Yes. Liminal Desktop has built-in dictation and text-to-speech. Speak a task and hear responses read back. The CLI web UI supports voice on the developer path too.
Can I run Liminal without my own API key?
On Pro, yes. Vireon-managed inference routes model calls through an OpenAI-compatible proxy with included monthly credits. No provider key required. Community Edition is bring-your-own-key and pays the provider directly with no markup.