Rasputin Mantle

Self-hosted autonomous agent platform.

Give it a goal. It browses, codes, researches, and shows you its work. You keep the keys.

Quick start · Architecture · Performance · v1.2 · Invariants · Gaps

---

Quick start (local dev)

git clone https://github.com/jcartu/rasputin-mantle.git
cd rasputin-mantle
cp .env.example .env
# fill in at minimum: ANTHROPIC_API_KEY, VLLM_BASE_URL, VLLM_MODEL

# Start vLLM separately on the host (e.g., port 8001 with qwen3.6-27b)

# Start the stack:
docker compose -f docker-compose.dev.yml up -d
docker compose -f docker-compose.dev.yml logs -f gateway web  # tail until ready

Open http://localhost:3000.

Ports:

• 3000 — web UI
• 8000 — gateway API
• 8025 — MailHog UI
• 8080 — sandbox / Neko
• 5432 — postgres (container-only)
• 6379 — redis (container-only)

Requirements: Docker, vLLM running on host (separate process), Anthropic API key for full feature parity.

---

Quick start (installer)

curl -sSL https://mantle.dev/install | sh

One line. The installer clones the repo, configures your API keys, brings up the Docker stack, and opens http://localhost:3000. Re-running is idempotent.

Requirements: Docker, Python ≥3.11, Node ≥18, pnpm, uv, and an LLM API key (Anthropic, OpenAI, or a local vLLM endpoint).

---

What this is

Rasputin Mantle is a self-hosted, MIT-licensed agent platform. The orchestration runs on your machine. The model spend lives in your account. No SaaS dependency, no billing layer, no opaque backend.

v1.2 ships a complete product surface: documented design system, 15 Radix-backed component primitives, three-pane resizable shell, real-time agent trace, Neko WebRTC live browser, mobile layouts, and full WCAG AA accessibility.

Underneath, the engine remains:

• FastAPI gateway with server-enforced cost ceiling (HTTP 429 on budget exceed)
• CodeAct executor (Wang et al. 2024) — Python action loop, open tool dialect
• Hardened Docker sandbox — python:3.12-slim, non-root, 512 MB RAM, no-new-privileges
• Playwright + Chromium — OS sandbox always on, enforced by pre-commit hook
• SKILL.md registry — frontmatter + body, loadable from any repo
• Neko WebRTC live view — real virtual browser stream, not screenshot polling
• rasputin-memory — Qdrant + FalkorDB, 72.40% LoCoMo
• Wide research — up to 10 parallel sandboxed subagents via asyncio.gather
• Voice loop — Faster-Whisper STT + Kokoro TTS, p50 365 ms
• MCP host — JSON-RPC 2.0, stdio + WebSocket
• Tauri 2 desktop — Mac DMG, Linux DEB/RPM, Windows MSI

---

Architecture

Five layers. Interface → Gateway → Agent Loop → Sandboxes → Foundation. Every layer is self-hostable. Cloud APIs are swappable backends, not requirements.

---

Performance

WebVoyager-300 — measured, reproducible

Same agent loop, same browser, same sandbox. The only variable is the planning model.

Eval	Result	Target
WebVoyager-300 (Opus 4.6)	71.00%	75%
Productivity skills	93.33%	80%
CodeAct contract	6/6 = 100%	6/6
Voice round-trip (p50)	365 ms	<1500 ms
Memory (LoCoMo)	72.40%	—
Lighthouse (landing)	97	95

Full eval artifacts in MANTLE_V1_1_RELEASED.md. Lighthouse scores and cost summary in MANTLE_V1_2_RELEASED.md. v1.3.2 benchmark deltas and methodology notes are in MANTLE_V1_3_2_RELEASED.md.

---

What ships in v1.3

• Onboarding flow — Three-step onboarding with intent selection and templates
• Playbooks gallery — Curated templates and user-saved workflows
• Empty state component — Reusable empty state pattern across the app
• Agent Skills standard — Anthropic-compatible SKILL.md directories with scripts, templates, examples, and discovery frontmatter
• Skills marketplace — Browse public, personal, and team skill tabs with install/use affordances plus save-session-as-skill support
• Projects — Persistent workspaces with project sidebar navigation, default planner/tool settings, and project-scoped sessions
• Knowledge base — Per-project file uploads capped at 50 files/100MB and mounted read-only into new project sessions
• Productivity skills — Bundled Slides, Spreadsheet, and Document skills produce real .pptx, .xlsx, .docx, and .pdf artifacts from structured outlines
• Productivity skills now use LLM-driven content generation — Slides, Spreadsheet, and Document skills generate content from facts and schemas instead of keyword matching
• Eval mode — MANTLE_EVAL_MODE=1 disables background scheduler, skill rediscovery, and knowledge-base lazy mounting for clean benchmark runs
• Artifact previews — Session artifacts lazy-preview PowerPoint first slides, Excel sheet samples, Word text, and PDF.js documents with downloads intact
• Productivity benchmark — 30 single-attempt tasks cover presentations, spreadsheets, and documents with structural judging and pass-rate reporting

What ships in v1.2

v1.0 proved the platform. v1.1 proved the agent. v1.2 makes both feel like a product.

Design system

A complete, documented design system in design-system/. Teal-sage accent. Geist Sans + IBM Plex Mono. Light and dark mode, both first-class.

• tokens.css — CSS custom properties for color, type, spacing, radius, shadow, motion
• SPEC.md — component contract. Every primitive, layout rule, interaction state
• COPY.md — voice and tone. Sentence case. Active voice. Forbidden words list
• MOTION.md — motion vocabulary. Standard easings, durations, reduced-motion

15 component primitives

Built on Radix UI. Styled against design tokens. Storybook stories and unit tests for each.

Button · Input · Tab · Dialog · Sheet · Textarea · Select · Switch · Tooltip · Toast · Badge · Avatar · Card · Skeleton · Spinner

Session view

Three resizable panes: agent trace (left), Neko WebRTC live browser (center), artifact list (right). Structured events streamed over SSE. Collapsible. Keyboard navigable.

Mobile layouts

All screens ship dedicated mobile breakpoints. Three-pane desktop becomes tabbed single-pane below 768px. Touch-aware Neko wrapper.

Accessibility

Full WCAG AA. Keyboard navigation for the complete task flow. Reduced-motion respected. Color contrast verified by Lighthouse across both themes.

---

Live Computer View

Most agent frameworks hide their browser. Mantle streams it back.

A real Neko WebRTC virtual browser lives inside the sandbox. When the agent navigates, clicks, types, or evaluates — you see it happen in real time. No screenshot polling. No fake animation. No SSE-throttled image strip.

• Trust by inspection. See what the agent sees. Intervene before destructive actions.
• Debugging by witness. Watch a task fail at step 7 of 10. No log archaeology.
• Portable proof. A SKILL.md author demos their skill; the WebRTC stream is the evidence.

---

Wide Research

from wide_research.dispatcher import dispatch_research

results = await dispatch_research(
    query="latest self-hosted LLM agent frameworks",
    n_agents=10,
    backends=["brave", "exa"],
)

Ten sandboxed subagents fan out in parallel via asyncio.gather. A merger collates, deduplicates, and produces a unified summary. Wall-clock time is roughly that of a single sequential search.

---

Engineering invariants

These rules the codebase will not violate. Enforced by pre-commit hooks, the auditor, or both.

1. Everything is self-hostable. No required SaaS.
2. License-clean. Every dep passes license-review before merge.
3. CodeAct sandboxing is non-negotiable. Agent never executes on the host.
4. Chromium sandbox is on. Always.
5. Cost ceiling is server-trusted. Middleware reads usage from upstream, writes to gateway_costs.
6. Reversible by default. State-mutating tools require confirm: true.
7. Gateway binds 127.0.0.1 unless MANTLE_PUBLIC=true.
8. Atomic state. Orchestrator survives kill -9. State persists via fsync + rename.

Sandbox hardening:

image        = "python:3.12-slim"
user         = "1000:1000"
mem_limit    = "512m"
cpu_quota    = 100_000  # ~1.0 CPU
security_opt = ["no-new-privileges:true"]
exec_timeout = 120

---

Honest gaps

• WebVoyager-300 — Canonical v1.3.2 result is 71% with Opus 4.6, improved from 67.67% in v1.3.1 but below the 75% target. The benchmark is hard. Kimi K2.6 at 69.33% is the floor on the -300 suite. (On the older WebVoyager-100, local Qwen3-235B bottomed out at 12%.)
• Productivity benchmark — Canonical v1.3.2 result is 93.33%, exceeding the 80% target under the facts-plus-schema methodology.
• Voice latency p95 — 2153 ms, elevated by Kokoro's first-iteration cold start. p50 of 365 ms is the steady-state number.
• STT on CPU — Faster-Whisper at int8 runs roughly real-time. Sub-100 ms STT requires a GPU.
• Tauri AppImage — DEB and RPM compile. AppImage bundling fails on icon manifest and is skipped.

Full audit log in AUDIT_2026_05_16.md. Every claim is reproducible.

---

On the roadmap (v1.3+)

• Replay mode — Completed sessions as a scrubbable timeline with playhead, tool-call jumps, and side-by-side screenshot/DOM diffs
• Share links — Signed share tokens for read-only replay, no cost data, expiry and optional passphrase
• Onboarding — Multi-step flow: API key entry → permission consent → first task suggestions, idempotent config writes
• Tauri desktop — Mac DMG, Linux DEB/RPM, Windows MSI binaries (config exists, binaries pending)
• Browser extension — Chrome/Edge "Send to Mantle" extension for daily-use surface
• Projects — Persistent workspaces with knowledge base, config inheritance, and session inheritance
• Skills standard — Anthropic Agent Skills format adoption, marketplace, one-click "save as skill"
• Integrations — Slack, email, scheduled tasks UI

---

Documentation

Document	Contents
docs/ARCHITECTURE.md	Component graph, data flow, request lifecycle
docs/SECURITY.md	Threat model, sandboxing, key handling, SSRF policy
docs/AGENT_SKILLS_SPEC.md	Anthropic Agent Skills-compatible directory format, frontmatter, discovery, and invocation
docs/SKILL_AUTHORING.md	Writing a SKILL.md, frontmatter schema, validation
MANTLE_V1_2_RELEASED.md	v1.2 release notes, Lighthouse scores, cost summary
MANTLE_V1_3_2_RELEASED.md	v1.3.2 release notes, benchmark deltas, methodology notes
CONTRIBUTING.md	Dev setup, coding style, commit format, PR flow

---

Reproduce locally

# 1. Clone with submodules
git clone --recursive https://github.com/jcartu/rasputin-mantle && cd rasputin-mantle

# 2. Configure
cp .env.example .env
$EDITOR .env   # ANTHROPIC_API_KEY, VLLM_BASE_URL, VLLM_API_KEY

# 3. Install
uv sync
pnpm install

# 4. Bring up full stack
docker compose -f infra/compose.dev.yml up -d

# 5. Verify
make verify

# 6. Run the web UI
pnpm --filter web dev   # http://127.0.0.1:3000

---

Citations

• CodeAct — Wang et al., Executable Code Actions Elicit Better LLM Agents, ICML 2024. arXiv:2402.01030
• WebVoyager — He et al., WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models, ACL 2024. arXiv:2401.13919
• LoCoMo — Maharana et al., Evaluating Very Long-Term Conversational Memory of LLM Agents, ACL 2024. arXiv:2402.17753
• Neko — m1k1o/neko — WebRTC virtual browser
• MCP — Model Context Protocol — JSON-RPC 2.0

---

License

MIT. Every dependency is OSI-approved. The license-review CI gate refuses non-compatible licenses.

---

_{Hand it a goal. Watch it work. Keep the keys.
Brand imagery generated with Gemini.}