Use SSH to expose project docs and agent workspaces through a shell-native filesystem, with a browser viewer for humans.
Docs: https://trknhr.github.io/docs-ssh/
Coding agents produce a lot of temporary knowledge:
- implementation plans
- investigation notes
- benchmark results
- task handoffs
This context may be useful for days or weeks, but not forever. Chat history can disappear, while committing every intermediate artifact to Git mixes short-lived notes with durable project documentation.
docs-ssh is a workspace between chat history and Git: persistent enough for agents and humans to revisit, but separate from the source repository. Agents explore it with familiar shell tools over SSH, while humans can use the browser viewer.
- Design and motivation: Project Docs over SSH for AI Agents
- MultiHop-RAG benchmark design and results
- Documentation
pnpm install
pnpm run build
npm link
pnpm run devDefaults:
- SSH server:
127.0.0.1:2222 - Viewer:
http://127.0.0.1:3000
From another terminal:
docs-ssh status --json
ssh localhost -p 2222 bootstrap --jsonpnpm test
pnpm run build
pnpm run site:build
pnpm run smokeUseful CLI flows:
docs-ssh config init
docs-ssh login --json
docs-ssh token login --token dssh_... --host docs-ssh --project default --json
docs-ssh skill --output .agents/skills/docs-ssh/SKILL.md- Runtime target is Node 24.
- Release images are published from
v*.*.*tags. - Use
pnpm run smokebefore release tags.
Git is still the right home for durable docs that should evolve with the code, such as READMEs, runbooks, and ADRs. docs-ssh is for intermediate work that is useful now but likely to become stale: discarded plans, investigation traces, verification results, and handoff notes.
docs-ssh does use SQLite for authentication and session metadata. Workspace content stays file-based so humans and agents can inspect it with ordinary tools without a database-specific query layer.
MCP is an integration protocol, not a storage policy. docs-ssh deliberately uses an interface that agents already understand: SSH plus filesystem commands. The two approaches are complementary; an MCP server could expose the same workspace.
Agents already know how to use ls, find, rg, and small file reads. SSH makes those operations remote, authenticated, scriptable, and easy to trace, without requiring a project-specific client. The browser viewer remains available for humans.