Silica, a notes/codebase agent that reverts its edits if they break coherence

Silica, an open-source agent for notes and codebases, uses a live model of the content to make LLM-driven edits that are automatically rolled back if they break coherence. The tool supports local-first operation, Obsidian integration, and multiple interfaces including a GUI, CLI, and MCP server.

Silica istantly knows what you're working on, helping you edit and organize your content without corrupting it. Point it at a folder of notes or a codebase. Silica builds a live model of what's inside, so the LLM answers from your actual material, and every edit it makes is checked and rolled back if it broke something. Local-first , Obsidian-compatible. presentation.mp4 License:AGPL-3.0-or-later,strong copyleft. Copying any part obliges your whole project to AGPL, including network-only use §13 . Details below . What you get what-you-get Quick Start quick-start Ways to use Silica ways-to-use-silica What you can do what-you-can-do See your vault see-your-vault How Silica actually works how-silica-actually-works Silica for codebases silica-for-codebases A promise a-promise Command reference command-reference Configuration configuration Status status References references Contributing contributing · License license A pile of pdf, or a repository, becomes something you can ask and reshape without babysitting it: Answers from what you actually have. Before you describe anything, Silica sees the shape of your vault: hubs, clusters, and notes nearest your questions so you get grounded answers. Edits don't rot in your vault. When Silica nucleates, merges, or refactors, every write can be reverted. The blast radius of a bad edit is one /undo . A graph you can navigate. Export the whole knowledge graph to an interactive page, or grow a radial mind-map out from any single note. Works offline. Local models LM Studio, Ollama are supported; if no embedder is configured, relatedness degrades to a deterministic local graph instead of failing. Still, embeddings are highly recommended for better quality. Clone the repository and install it in editable mode: git clone https://github.com/kiycoh/silica-agent.git cd silica-agent uv pip install -e . Optional features are installed as extras, alone or combined '. gui,mcp ' : uv pip install -e '. gui ' web GUI: silica --gui uv pip install -e '. mcp ' MCP server: silica mcp Silica as agent memory uv pip install -e '. connect ' Obsidian plugin bridge: silica connect uv pip install -e '. pdf ' PDF nucleation uv pip install -e '. dev ' tests and linters Run the interactive wizard to set up your .env vault, backend, chat provider, embeddings : uv run silica init Re-check the environment at any time: uv run silica doctor Start the interactive REPL: uv run silica A good first move on an existing vault is a read-only structural audit. It never writes, and it shows you the hubs, bridges, and orphans before you touch anything: /report The same vault model serves four different drivers. What changes is who holds the write key, and whether they read or write: - GUI A chat-first web interface default silica --gui http://localhost:8765 . Query and curate the vault from the browser, watch answers stream in, and open the graph. The friendliest way in, and the best first impression. - CLI / TUI The interactive terminal REPL. Every command in the silica reference command-reference lives here: nucleate, audit, search, refactor, visualize. Fastest for real work once you know the verbs. - Obsidian plugin A live bridge into the Obsidian desktop app, so Silica reads and writes the vault you already have open, with its rollback and cache backing every change. silica connect Feature-complete, pending end-to-end hardening. - MCP server Silica serves the vault over stdio to any MCP client. An assistant recalls from your real notes before it answers, grounding on your real decisions instead of guessing. For Claude Code, the repo is also a plugin: silica mcp claude plugin marketplace add /path/to/silica-agent claude plugin install silica@silica Clear inbox files without losing anything. Drop raw clippings and drafts in a folder; /nucleate Inbox/ .md distills each into an atomic note, checks it against what you already have so you don't get a fifth copy of the same idea, and files it. Hand it 20 files at once; it never gets confused by the pile. Ask your notes, not your memory. /explain "<concept " , /compare "A" "B" , /summarize <folder , /quiz <note : read-only answers grounded in the vault, with contradictions surfaced instead of smoothed over. Reorganize by intent. /organize "group by project" classifies and moves notes to a taxonomy. /curate plans the autolink, dedup, and cleanup work; add --apply to run it. Refactor safely. Merges and splits redirect every incoming link automatically, so you never end up with a broken reference or an orphan. Changed your mind? /undo a note or /revert a whole run. Research into the vault. /web-search "<topic " pulls cited findings into the inbox; /nucleate them when you're ready. Nothing from the web lands in the vault without your say-so. Silica turns the invisible structure of your notes into something on screen. Knowledge graph /graph out.html exports an interactive page: notes as nodes, links as edges, communities colored and named automatically so the clusters read at a glance. Open it in any browser, no server needed. Mind-map /map <note grows a radial map rooted on one note and writes it as maps/<stem .canvas opens natively in Obsidian plus an SVG. Layout is by community, laid out so nodes never overlap, so the picture is legible from the first render.Both run locally and work embedder-free: even with no embedding model, the deterministic co-occurrence graph still gives you clusters and relatedness. Can be viewed in GUI or saved as file. You already let deterministic tools rewrite and reject your work every day. You don't extend them trust; you trust the guardrail. Silica wraps an LLM's edits to your vault in the same kind of guardrail: | You already let a deterministic tool… | to guard against… | Silica does the same for a vault by… | |---|---|---| a compiler reject source that won't build | syntax and type errors | an FSM refusing to commit a note that fails its structural checks | a test suite block a merge that breaks behavior | regressions | a post-write verify gate that reverts any edit which breaks vault coherence | git roll back a bad commit | losing history | /undo and /revert rolling back an injection, per-note or per-run | a formatter rewrite your code without asking | drift and inconsistency | graph-safe refactors that redirect links so a merge or split never orphans a note | Silica is not a free-form agent. Every vault mutation passes through a finite-state machine that enforces these contracts: Single entry point: all nucleation flows through the Injector FSM. There is no side channel that writes to the vault. Verify-or-revert: every write is re-read and checked afterward; a mismatch VerifyMismatchError rolls the write back. Graph-safe moves: renames, merges, and splits redirect incoming links atomically. No operation leaves a broken reference or an orphan. Zero-trust ingress: external content e.g. web search can only land in Inbox/ . Nothing reaches the vault without explicit human staging and FSM review. Layered rollback: /undo per note , /revert per run , and optional SILICA GIT COMMIT=auto stack as independent safety nets. Point SILICA VAULT at a repository instead of a note folder, and Silica keeps a living, human-readable map of the code under docs/silica/ , honest against git. /nucleate <file extracts a shallow AST skeleton with tree-sitter signatures, structure, imports and turns it into a markdown note that documents its source, stamped with the commit it was verified against. /wiki grows that into a behavioral wiki: an ARCHITECTURE.md plus one note per subsystem. /stale flags notes whose source changed in shape since you last documented it a signature or control-flow change, not just a reformat ; /impact maps a set of changed files to the notes they affect. You re-document what actually moved, not the whole tree. Two readers, one artifact: a human reads it as an always-current map, and a coding agent reads it over the MCP server ways-to-use-silica to ground its work in the real structure instead of re-deriving it every session. Most "chat with your notes" tools hand an open-ended agent a write key and hope the edits are good. The failure is quiet: a merge orphans a note, a rewrite breaks a wikilink. There is no free-running loop holding a key to your vault. Every change goes one way: the agent proposes, a state machine checks it, you confirm, and the write is re-read and rolled back if it broke coherence. You can still use git alongside Silica as the byte-level backstop; Silica is the coherence layer on top. It is always highly recommended to create a backup of your vault before allowing Silica to manipulate your files.⚠️ Note: That's the whole idea. You don't have to believe anything about the model, only that the guardrail runs on every write. It does. Enforced today on the normal write path. Not yet crash-verified a harness that kills the process mid-write to prove invariants survive failure is in progress : trust it as enforced control flow, not a proof under adversarial faults. Ask & audit read-only : | Command | What it does | |---|---| /report folder | Structural audit: hubs, bridges, orphans | /explain "<concept " --level | Explain a concept, grounded in the vault | /summarize <note|folder | Digest of one or more notes | /compare "A" "B" | Comparison table, surfaces contradictions | /quiz <note --n=10 | Active-recall quiz from your notes | /relate <note · /path A B | How notes relate · shortest reading path | /find <query | Semantic search | Bring in & reshape: | Command | What it does | |---|---| /nucleate <file... --target=DIR | Notes via the gate; code as skeletons | /organize "<intent " --apply | Classify and move notes to a taxonomy | /curate --apply · /dedup · /refine | Plan / run autolink, dedup, enrichment | /web-search "<topic " | Cited web findings into the inbox | /convert <file | Transcode a PDF into a markdown draft | Visualize: /graph out.html · /map <note Codebase: /wiki · /stale · /impact <range Undo & inspect: /undo note · /revert run · /status · /review System: /help · /model · /vault path · /settings <key <value · /clear · /exit silica init writes the essentials; the full list with defaults is in .env.example /kiycoh/silica-agent/blob/main/.env.example . | Variable | Description | |---|---| SILICA MODEL | Chat model litellm format, e.g. openrouter/anthropic/claude-sonnet-4 | SILICA PROVIDER | lmstudio or openrouter | SILICA VAULT | Vault path. Obsidian vault used verbatim; any other path is repo mode → docs/silica/ | SILICA EMBEDDING MODEL | Embedding model for semantic tasks default qwen3-embedding-4b | SILICA BACKEND | fs default, headless or cli live Obsidian via CDP | SILICA GIT COMMIT | Git safety net for writes off , auto | SILICA TAVILY API KEY | Enables /web-search | SILICA WORKER MODEL | Sub-agent worker model for dedup/refinement operations | Available now: note nucleation, structural audit /report , semantic and embedder-free search, graph-safe refactor / dedup / merge, graph and mind-map export, codebase skeletons with git-backed /stale and /impact , the code /wiki , layered /undo and /revert , git safety net, the MCP server and Claude Code plugin. In progress: richer codebase coverage across more languages, PDF/DOCX/TXT nucleation, the live Obsidian bridge silica connect , and the crash harness backing the guardrail. Planned: image nucleation, MCP packaging for non-Claude agents. What ships today is enforced; what's in-progress or planned is not yet present. arXiv:2604.11378, 2026 From Agent Loops to Structured Graphs: A Scheduler-Theoretic Framework for LLM Agent Execution https://arxiv.org/abs/2604.11378 arXiv:2606.11688, 2026 Goal-Autopilot: A Verifiable Anti-Fabrication Firewall for Unattended Long-Horizon Agents https://arxiv.org/abs/2606.11688 arXiv:2606.14831, 2026 Is Your Agent Playing Dead? Deployed LLM Agents Exhibit Constraint-Evasive Fabrication and Thanatosis https://arxiv.org/abs/2606.14831 arXiv:2601.08773, 2026 Reliable Graph-RAG for Codebases: AST-Derived Graphs vs LLM-Extracted Knowledge Graphs https://arxiv.org/abs/2601.08773 Predicting new research directions in materials science using large language models and concept graphs https://doi.org/10.1038/s42256-026-01206-y Nature Machine Intelligence , 2026 Silica's embedder-free near-duplicate detection /dedup is inspired by and ports the well-thought-out MinHash design from Graphify https://github.com/safishamsi/graphify . Issues and pull requests welcome. See CONTRIBUTING.md /kiycoh/silica-agent/blob/main/CONTRIBUTING.md for setup and conventions English-only, conventional commits . By contributing you license your work under AGPL-3.0-or-later. For security issues follow SECURITY.md /kiycoh/silica-agent/blob/main/SECURITY.md ; do not open a public issue. GNU Affero General Public License v3.0 AGPL-3.0-or-later . Copyright C 2026 Alessandro Carosia. Strong copyleft: incorporate any portion of Silica and that work becomes a derivative that must itself be AGPL-3.0, with complete corresponding source offered to everyone who uses it. §13 extends this to network use: running a modified version as a hosted service obliges you to provide source to your users. There is no permissive fallback. See LICENSE /kiycoh/silica-agent/blob/main/LICENSE for the full text.