# ERIS: Local-first vault agent in Rust, grammar-enforced tools

> Source: <https://github.com/janpauldahlke/eris>
> Published: 2026-09-04 09:05:04+00:00

**A local-first agent in a single Rust binary: your Markdown vault as memory, grammar-enforced tool calls on llama.cpp, and nothing leaves your machine unless you say so.**

**Eris** runs a local LLM as a personal agent over a plain-Markdown vault (Obsidian-compatible). It reads and writes your notes, remembers across sessions through tiered semantic memory, manages reminders and alarms, and calls its tools through a JSON protocol that is **structurally enforced by a GBNF grammar** — no function-calling API required, and no cloud in the loop.

Named for [Eris](https://en.wikipedia.org/wiki/Eris_(mythology)) — the Greek goddess of strife and the golden apple of discord. The name is mythological, not an acronym: a local agent meant to productively disturb your vault (read, write, remember, remind).

**Sovereign by architecture, not by promise.** No telemetry leaves the machine. Every outbound integration (web fetch, mail, calendar, Discord) is opt-in and enumerable. The vault is plain text you own; the semantic index is derived data that can always be rebuilt from it.**Reliability engineering for small models.** Local 8–26B models are sloppy with JSON. Eris compiles a GBNF grammar per session — and a narrowed per-turn subset grammar — so the model's output is constrained at the token level. Bounded generation (`n_predict_max`

) guards against context-window truncation.**One brain, three faces.** The same orchestrator and tool registry serve a full-screen terminal UI (ratatui), a localhost web UI (`eris chat --web`

), and an optional Discord sidecar sharing the live session.

Eris is **not** smarter than hosted frontier models — its intelligence ceiling is the GGUF you run. It is smarter *about your data* than anything hosted: it lives where your notes live, and it works when your network doesn't.

| Core (supported) | Extras (best-effort) |
|---|---|
| Chat (TUI + web), Markdown vault read/write/search | Discord sidecar |
| Tiered memory: staged (ephemeral) + committed (vault) + semantic recall (Qdrant) | Vision (`vision:see` , multimodal GGUF + mmproj) |
| Tool protocol with gatekeeper + GBNF enforcement (llama.cpp) | Voice ingress (STT via ffmpeg) |
| Agenda: reminders, alarms, self-driven follow-ups | Google Workspace mail/calendar tools |
| Web fetch/search with allowlist, consent, and session budget | Moltbook client |

Full tool roster (by family): ** docs/TOOLS.md**.

Prerequisites: Rust (stable, edition 2024), [llama.cpp](https://github.com/ggml-org/llama.cpp) (`llama-server`

), a chat GGUF + an embedding GGUF (e.g. `nomic-embed-text`

), and Qdrant for semantic memory:

```
docker run -d -p 6333:6333 -p 6334:6334 -v eris-qdrant-data:/qdrant/storage qdrant/qdrant
```

Build and run:

```
cargo build --release
./target/release/eris chat          # first run launches the ignition wizard
./target/release/eris chat --web    # same session, localhost web UI
```

The first-run wizard writes `.fcp/config.toml`

(backend, model paths, GPU layers) and seals the vault directory. Full setup — including the Ollama alternative backend, vision, and voice — is in ** docs/REFERENCE.md** and

**.**

[docs/HOW_TO/](/janpauldahlke/eris/blob/main/docs/HOW_TO)| Setup | Works |
|---|---|
| Apple Silicon, 16 GB+ | Good: 7–12B GGUF chat + embed model, Metal offload |
| Apple Silicon, 32 GB+ | Comfortable: 26B-class models, vision mmproj |
| Linux + NVIDIA (8 GB+ VRAM) | Good with `--n-gpu-layers` tuning |
| CPU-only | Runs, but slow; small quantized models only |

**llama.cpp is the canonical production backend** — it is the only one with GBNF grammar enforcement, vision, and voice. Ollama is supported as an easier-to-install alternative with weaker JSON discipline (soft `format: json`

instead of grammar); expect more recovery turns on long sessions.

Alpha. Single-user, single-process. Dogfooded on **macOS** (Apple Silicon / unified memory, Metal) and **Linux** (ideally NVIDIA GPU). Windows should run; it is less exercised and packaging is not a priority yet. Honest known limitations:

- Long-context sessions can still degrade JSON discipline on the Ollama backend (grammar-less path).
- Installation is manual (build from source + fetch models); installers and prebuilt binaries are planned.
- The architecture docs in
[docs/updated_architecture/](/janpauldahlke/eris/blob/main/docs/updated_architecture)include a frank[self-review](/janpauldahlke/eris/blob/main/docs/updated_architecture/10_DEEP_REVIEW_2026-07.md)of the codebase's debt — read it before contributing to the orchestrator.

Everything under [ docs/](/janpauldahlke/eris/blob/main/docs) is

**my working notes**— design diaries, reviews, how-tos, and TODOs I keep in-repo so the thinking stays visible and I can sync across machines. I curate them when I can; they are

**not** a guarantee of being fully up to date with

`main`

. Prefer the code, this README, [, and](/janpauldahlke/eris/blob/main/SECURITY.md)

`SECURITY.md`

[when something conflicts.](/janpauldahlke/eris/blob/main/CONTRIBUTING.md)

`CONTRIBUTING.md`

| Doc | Contents |
|---|---|
|

[docs/REFERENCE.md](/janpauldahlke/eris/blob/main/docs/REFERENCE.md)[docs/HOW_TO/](/janpauldahlke/eris/blob/main/docs/HOW_TO)[docs/updated_architecture/](/janpauldahlke/eris/blob/main/docs/updated_architecture/README.md)Contributions are welcome under **inbound = outbound** terms: your contributions are licensed under Apache 2.0, confirmed by a [DCO](https://developercertificate.org/) `Signed-off-by`

line (`git commit -s`

). See [CONTRIBUTING.md](/janpauldahlke/eris/blob/main/CONTRIBUTING.md) — including the project's non-negotiable engineering rules (zero panics, no `unsafe`

, actor-model concurrency).

Copyright 2026 Jan Dahlke. Licensed under the [Apache License, Version 2.0](/janpauldahlke/eris/blob/main/LICENSE).