Use Scribe with a local llama.cpp, Ollama, or LM Studio server, or point it at a cloud API on weaker hardware. It keeps sessions, files, RAG, and memory in plain local folders — and in 3.0 its three guarantees are measured: a public leaderboard ranks models on citation discipline, and tool calls survive on every backend.
$ git clone https://github.com/pedjaurosevic/scribe-llm.git
$ cd scribe-llm && ./scripts/install.sh
$ scribe-llm config show
base_url: http://127.0.0.1:18083/v1
workspace: ~/scribe-workspace
$ scribe-llm chat
# Work with files, sessions, RAG, and tools
# inside your local Scribe workspace.Three properties that hold by construction, not by prompt-tuning — and an instrument that puts a number on them. Full story in the whitepaper.
On llama.cpp a GBNF grammar makes a malformed call grammatically impossible. New in 3.0: everywhere else the same forced-call path degrades to structured outputs (json_schema → json_object → text parse).
Grounded answers map every claim to a numbered source [n], tag contradictions, and refuse outside the sources — even when the model "knows" the answer.
The held-out suite was hardened in 3.0 (8 → 24 tasks) after the old one saturated. Deterministic SPI, checksum-locked tasks, no LLM judge.
Seven models, one identical grounded prompt, 24 locked tasks. Local 4–12B models match or beat 70B/120B cloud models on citation discipline. Full report · reproduce with scribe-llm bench --models.
The GitHub install is editable, so a later git pull updates the same checkout. The install script also creates ~/.config/scribe/config.toml and ~/scribe-workspace.
git clone https://github.com/pedjaurosevic/scribe-llm.git
cd scribe-llm
./scripts/install.sh
scribe-llm --versionpython3 -m pip install scribe-llm
scribe-llm --help
scribe-llm chatScribe talks to OpenAI-compatible APIs. On llama.cpp it can use grammar-constrained tool calls; on cloud providers it falls back to regular tool-call/text parsing.
Good for private work, local files, and long-context setups. Default endpoint is http://127.0.0.1:18083/v1.
Use a smaller model locally, or test Scribe without running a custom server. Set base_url and model.
Useful on weaker machines. Keep the API key in SCRIBE_API_KEY, not in a committed config file.
[scribe]
base_url = "https://api.deepseek.com"
model = "your-deepseek-model"
api_key = "use-env-instead"
reasoning = false
tool_grammar = "off"export SCRIBE_BASE_URL="https://api.deepseek.com"
export SCRIBE_MODEL="your-deepseek-model"
export SCRIBE_API_KEY="..."
scribe-llm chatThe project is a practical agent stack: model adapter, session manager, local tools, RAG, semantic memory, and a web writing studio.
The agent reads and writes inside a configured workspace directory, with scoped file operations.
Web search, web fetch, source-grounded answers, contradiction tags, and citation-oriented prompts.
SQLite FTS5 plus vector search with multilingual embeddings, combined with reciprocal rank fusion.
Cross-session recall through SME and a durable WorldModel for persistent context.
Command gates, Python AST checks, sandbox support, and git checkpoints for safer local changes.
A localhost web editor for chapters, outlines, terminal work, and Markdown/PDF/EPUB export.
Scribe 3.0 ships as stable on PyPI, tested by a 361-test suite (10 tests new in 3.0, covering the structured-outputs fallback and leaderboard robustness) and an 18/18 live gauntlet. The site should still make the boundaries clear instead of overselling them.
base_url, model, and api_key.json_schema → json_object → text parse) instead of best-effort parsing alone.scribe-llm chatinteractive terminal chatscribe-llm weblocalhost Book Studioscribe-llm discoverfind local model serversscribe-llm status --jsonmachine-readable statusscribe-llm rag searchsearch indexed documentsscribe-llm initcreate a project vaultscribe-llm compareblind A/B model testscribe-llm benchquality and grounding checks