Get This Tool
Scribe
Pricing
- Model
- Free
Summary
Agent sessions that don't know what you decided last week repeat your worst mistakes — same library choices, same architectural debates, zero institutional memory. Scribe exists to fix that: a cron-driven Go binary that reads your git history and coding-agent sessions, then writes a cross-project wiki so the next agent already knows what you settled and why.
The pipeline runs in three stages — capture from git repos, Claude Code/Codex sessions, iMessage-dropped URLs, and drop files; triage via keyword-density scoring before any LLM call touches the input; then compile into a typed-graph wiki of plain markdown with auto-generated wikilinks and backlinks. FTS5 pre-filtering means sessions with nothing worth keeping cost zero API tokens. The vendor states the full pipeline runs against a local Ollama server with one line changed in scribe.yaml, eliminating API spend entirely. Where it strains: the system is only as useful as the agent's ability to query the KB — if your agent doesn't read CLAUDE.md or AGENTS.md on startup, the handshake Scribe writes is ignored. Teams adopting this commit to a local-first, terminal-native workflow; there is no hosted dashboard or GUI to inspect what Scribe wrote.
Bottom line: Pick Scribe when you want agents to stop re-litigating architectural decisions across projects and you're willing to run a cron daemon locally — but if your team needs a shared, access-controlled knowledge base across more than one developer's machine, the git-native single-user model forces you to bolt on your own sync layer.
Community Performance Report Card
No community ratings yet. Be the first to rate this tool!
Community Benchmarks Community
Sign in to submit a benchmarkNo community benchmarks yet. Be the first to share a real-world data point.
Pros
Sign in to edit- FTS5 pre-filtering rejects low-value sessions before any LLM call, so zero-insight sessions cost zero tokens — which means running the pipeline daily against a local Ollama server produces no API bill and no noise in the output corpus.
- The CLAUDE.md and AGENTS.md handshake injection means every new agent session in every project queries prior decisions automatically, so agents stop proposing the library you already evaluated and rejected two projects ago.
- A typed 10-kind edge schema (supersedes, contradicts, derived_from, and others) connects articles relationally, so the KB captures decision lineage rather than just a flat pile of notes that have no memory of which choice replaced which.
- One flag in scribe.yaml switches the entire pipeline to a local Ollama server, which means you can run extraction, absorb, and Dream consolidation with zero external API dependency and no data leaving your machine.
- Output is plain markdown in a git repo you control, so the corpus survives any change to Scribe itself — open it in Obsidian, VS Code, or vim, push it to Gitea or Forgejo, and no vendor has leverage over your data.
Cons
Sign in to edit- The handshake only works if the agent reads CLAUDE.md or AGENTS.md at session start — agents or wrappers that skip that file ignore every decision Scribe wrote, and there is no fallback mechanism to force context injection, so teams with non-standard agent setups get no benefit from the KB without manual wiring.
- Scribe is a single-developer local tool with no multi-user access model: one machine, one corpus, one git repo. Teams of two or more who need shared architectural memory hit this wall immediately and end up running separate personal instances while syncing decisions by hand — at which point they are maintaining the problem Scribe was supposed to solve.
- The Dream consolidation cycle runs on a fixed cron schedule (weekly, Sundays at 02:00 per the docs), so a decision made Monday that contradicts a decision made Friday won't be cross-linked until the next consolidation window — teams working fast on interrelated projects find the KB is always one cycle behind the work.
- There is no GUI, no hosted dashboard, and no search UI outside the terminal qmd command or Claude Code's MCP integration — teams whose product managers or non-engineering stakeholders need to read the KB have no access path without standing up a separate rendering layer like mdbook or Obsidian.
Community Reviews
Sign in to write a reviewNo reviews yet. Be the first to share your experience.
About
- Platforms
- macOS, Linux (Go binary)
- API Available
- No
- Self-Hosted
- Yes
- Last Updated
- 2026-07-19T20:23:11.152Z
Best For
Who it's for
- Developers using Claude Code or Codex who want agents to reference prior decisions
- Users wanting a fully local, git-native KB with typed relations
- Teams needing cron-driven consolidation of work across projects
What it does well
- Auto-maintain cross-project decision history for AI coding agents
- Capture and link architectural choices across multiple repositories
- Build a searchable plain-markdown knowledge base without manual note-taking
Integrations
Discussion Community
Sign in to commentNo discussion yet. Sign in to start the conversation.
Compare Scribe
Spotted incorrect or missing data? Join our community of contributors.
Sign Up to ContributeCommunity Notes & Tips Community
Sign in to contributeBe the first to contribute. General notes, observations, gotchas, and tips from people who use this tool day-to-day.
Frequently Asked Questions
- Is Scribe free?
- Yes — Scribe is fully free to use. There is no paid tier.
- Is Scribe open source?
- Yes. Scribe is open source.
- Can I self-host Scribe?
- Yes. Scribe supports self-hosting on your own infrastructure.
- What platforms does Scribe support?
- Scribe is available on: macOS, Linux (Go binary).
Hours Saved & ROI Stories Community
Sign in to contributeBe the first to contribute. Concrete time/cost savings, with context. e.g. "Cut my code review backlog from 4h to 45m per week."
Curated lists that include this category
Scribe is a single Go binary that auto-maintains a cross-project knowledge base by watching four input streams — git commit history, Claude Code and Codex session logs, URLs texted to yourself via iMessage, and drop files from other projects. A cron schedule drives the whole pipeline: every two hours it extracts per-project decisions, three times a day it mines agent sessions, and a weekly ‘Dream’ consolidation pass compresses and cross-links accumulated knowledge. The output is plain markdown with YAML frontmatter, stored in a git repo you own and push wherever you want.
The differentiating mechanic is the two-pass absorb: pass one extracts atomic facts, pass two fans dense sources into multiple entity-first wiki pages. Before any LLM processes a session, an FTS5 keyword-density scorer rejects boilerplate — so sessions with nothing extractable cost nothing, whether you’re on Ollama or an external API. Articles connect through a closed 10-kind typed-edge schema (supersedes, contradicts, derived_from, specializes, extends, and five others), which means the graph captures not just what you decided but how decisions relate — something flat tags cannot express.
Scribe fits developers who run Claude Code or Codex across multiple repos and are tired of re-explaining the same context at the start of every session. The scribe init command writes a handshake block into both ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md, so agents query the KB before proposing any library or architecture. Where Scribe doesn’t fit: it is a single-developer, local-first tool. There is no multi-user access model, no hosted sync, and no GUI — the corpus lives in a git repo you manage. Teams expecting a shared internal wiki for more than one engineer hit that ceiling quickly and end up treating Scribe as a personal layer while managing team knowledge elsewhere.