seraph
Health Warn
- License — License: MIT
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Low visibility — Only 5 GitHub stars
Code Pass
- Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Pass
- Permissions — No dangerous permissions requested
No AI report is available for this listing yet.
An AI guardian that remembers, watches, and acts.
Seraph
Your goals. A guardian that grows with you.
Building a local-first, self-improving AI guardian.
Website · Quick start · Documentation · Vision · Community
Seraph brings goals, permitted context, research, tasks, and memory into one
workspace. Give useful work a place to continue: review opportunities in your
Inbox, follow progress on the Work board, and help Seraph learn what helps you.
Watch the workspace demo
Recorded on an earlier cockpit build. See the Current App Guide for today's interface.
Seraph Today
Status: Partial — usable workflows, with the broader guardian vision still in development.
| What you can do | |
|---|---|
| 🎯 Keep goals in view | Bring goals, tasks, conversations, and evidence together; inspect progress and recover interrupted work. |
| 🔎 Find relevant signals | Assess changes from public source watches you select. Review cited opportunities in the Inbox, with attention and spending limits. |
| 🛠️ Turn evidence into work | Preview and accept bounded research, browser checks, reports, repository repair, and selected document workflows. |
| 🌱 Teach it what helps | Give explicit Helpful / Not helpful feedback, review preference recommendations, and separately adopt or roll back memory changes. |
| 🧩 Extend your workspace | Use skills, MCP adapters, reviewed procedures, and capability packages, plus optional consented desktop context and integrations. |
Available workflows depend on permissions, configuration, and host support.
See Development Status for exact limits and
v2026.11.17 release notes for the
latest test determinism, daemon authentication, screenshot cap and GitHub activity changes.
Quick Start
Python 3.12 · uv · Node.js 24 · npm
The core runs on a CPU host. No GPU or local model server is required.
For a fresh development checkout:
git clone https://github.com/seraph-quest/seraph.git
cd seraph
uv sync --project backend --locked --group dev
npm ci --prefix frontend
npm ci --prefix runtime/cordis --ignore-scripts
cp env.dev.example .env.dev
./manage.sh -e dev local up
./manage.sh -e dev local status
Open the cockpit or
API documentation. The core starts without a provider
key. To turn chat on, open Settings → Model: the Model card at the top says
whether chat works and shows the one control that fixes it (a NEAR API key, a
model, or your explicit OK that NEAR sees messages as plain text). Seraph sets
no spending limit of its own: set limits in your NEAR account. Vision and embedding have separate controls under Advanced and
diagnostics on the same tab.
Upgrade note: chat now runs on the Cordis agent loop by default, so chat
requires the Cordis runtime (Docker images and ./manage.sh -e dev local up
already build it). On the first start after upgrading, Seraph moves chat to the
Cordis loop once; if the runtime is missing, chat shows as blocked with the exact
fix instead of falling back. Settings → Model → Advanced and diagnostics → Cordis
runtime host can roll chat back to
the legacy loop (and cut it over again); a rollback is never undone by a restart.
Goal discovery (programme research) runs through the reviewed Cordis host. It
needs Node 22.12+ or 24.x (npm 11.8.0, bundled with Node 24.13.1): put it on PATH
or set CORDIS_NODE_PATH in .env.dev. After the npm ci above, local up
builds runtime/cordis when needed (it never installs dependencies) and the
backend starts the host. Without it, core startup continues and
discovery shows as blocked with the exact fix in Settings and the discovery
panel; ./manage.sh -e dev cordis build repeats the build.
Use ./manage.sh -e dev local run for a foreground session, particularly in
managed development shells. Stop with ./manage.sh -e dev local down.
Preserve your existing .env.dev when returning to an installed workspace.
Local-first means Seraph owns your canonical goals, memory, jobs, artifacts,
and history. It does not mean inference is offline. Ordinary text, vision,
and embedding use NEAR TEE-hosted models (per NEAR's catalog, not verified by
Seraph) under ADR-035; NEAR receives requests in plaintext and OpenRouter is
retired. The optional NEAR HTTPS text
capability is disabled by default, requires separate setup and consent, and
sends provider-readable plaintext.
The managed launcher owns environment loading and process lifecycle.
This quick start is for local development; see the
Current App Guide for
authentication, optional adapters, recovery, and production requirements.
The Self-Improving Guardian Vision
Seraph should feel like a guardian angel: understand your goals, priorities,
constraints, and progress; connect permitted context with relevant developments;
and prepare or perform useful work before you ask, within the authority you give it.
Imagine a speaking opportunity that fits your talk goal, a paper that unblocks
what you are learning, or a project obstacle caught early with a practical
response ready to review. The ambition is useful initiative across days—with
less setup, repeated explanation, and unnecessary interruption.
Self-improving means learning under your control. Explicit feedback and
verified outcomes should improve suggestions, procedures, and capabilities
through reviewed, reversible changes. Today's learning is bounded; generalized
source discovery, open-ended follow-through, and measured improvement across
unfamiliar tasks remain ambitions. Success alone does not authorize learning
or broader permissions.
The Project Constitution
defines this direction and its accepted boundaries.
- Choose a Goal, select relevant public source watches, and enable a finite
guardian policy with attention and spending limits. - Inspect a cited opportunity when a qualifying change arrives. Low-confidence
or irrelevant assessments stay quiet. - Preview and accept a supported browser-check or evidence-report plan in Work.
Review any required approval, then inspect the result and its sources. - Mark the opportunity Helpful or Not helpful. Review any resulting preference
separately before adopting it into memory; you can reverse it later.
The current loop starts with sources you choose. Relevance alone does not
authorize execution, and delivery does not count as positive feedback.
What We're Building Toward
- Turn goals into prioritized plans, scheduled work, and evidence-backed progress reviews.
- Monitor consented desktop context and produce searchable summaries that help the operator reflect and recover focus.
- Continuously research operator-selected topics and connect material findings to active goals and decisions.
- Execute bounded software-engineering and knowledge workflows with approvals, artifacts, checkpoints, and audit receipts.
- Continue one trusted conversation across the cockpit, paired voice, and paired messaging surfaces.
Explore and Contribute
| For | Start here |
|---|---|
| Running Seraph | Current App Guide · Support |
| Understanding the project | Development Status · Project Constitution |
| Building with us | Contributing · Agent guidelines · Documentation Contract |
| Questions and ideas | GitHub Discussions · Issues |
The current runtime is Python/FastAPI with a React cockpit. The accepted
all-plugin Cordis architecture
is Planned migration work.
Contributions that make the guardian more useful are welcome. Track substantial
work, use a feature or fix branch, validate the behavior, and obtain independent
review before merging. Report vulnerabilities through the Security policy.
Released under the MIT License.
Reviews (0)
Sign in to leave a review.
Leave a reviewNo results found