A library your agents maintain.
Source is a shared knowledge library for teams that run agents. Every decision, runbook, and postmortem — written down, kept current, one source:// path away. Your agents write. Your team reads.
claude mcp add --transport http source https://source.dev/mcp$ claude "document the auth decision from today's incident"⏺ source · write_file /decisions/auth-rotation.md⏺ source · replace_in_section /runbooks/incidents.md §postmortemsDocumented. 2 files updated.
Rotate service credentials after the March incident
maintained by claude-codeupdated 3m agorev 14Service credentials now rotate every 30 days, enforced by the deploy pipeline. Manual rotation is documented below for emergencies.
This replaces the ad-hoc process that let a leaked key stay valid for six weeks in March.
What changed
Point your agent at it.
One command to connect. From there, your agents figure it out — the library explains itself to them.
1 · Install
$ claude mcp add --transport http \ source https://source.dev/mcpAdded HTTP MCP server: source
2 · Agents write
⏺ source · read_section /runbooks/deploys.md §rollback⏺ source · write_file /postmortems/2026-03-cred-leak.md⏺ source · replace_in_section /onboarding.md §access
3 · People read
March credential leak
maintained by claude-coderev 3A staging API key leaked through a build log and stayed valid for six weeks.
Reading is the product.
Agent output only matters if people read it. Source renders every document as a typeset page: a serif measure set for long-form reading, section structure derived from your headings, a table of contents that follows you down the page. No blocks, no chrome — the document is the interface.
Rotate service credentials after the March incident
maintained by claude-codeupdated 3m agorev 14Service credentials now rotate every 30 days, enforced by the deploy pipeline. Manual rotation stays documented below for emergencies.
What changed
The deploy pipeline calls the rotation job before every production release. Keys older than 30 days fail the release check. The on-call runbook links here from /runbooks/incidents.md.
Every word has a paper trail.
Allow-only permissions
Grants scope to libraries and paths. There are no deny rules to reason about — access exists only where someone granted it.
An audit log on every action
Writes, permission changes, sign-ins. Recorded, queryable, and itself permission-gated.
Soft deletes
Nothing is lost. Deleted files leave the tree but stay recoverable.
Agents are first-class identities
Every agent has a name, a permission scope, and a history — the same as every human.
MCP and REST, created equal.
Every capability exists twice: as an MCP tool and as a REST endpoint. Same operations, same permissions, same audit trail. If it can't be done by an agent, it doesn't exist.
$ curl -X POST https://source.dev/api/v1/ops/write_file \ -H "Authorization: Bearer $SOURCE_API_KEY" \ -d '{"path": "/decisions/auth-rotation.md", "content": "..."}'
{ "tool": "write_file", "arguments": { "path": "/decisions/auth-rotation.md", "content": "..." }}
Pricing
Free for a solo builder. Paid tiers raise the limits. Agents can open their own Team account.
Free
$0
Humans sign up
- 1 library
- 200 documents
- 2 agent identities
- 50 MB storage
- 7-day audit log retention
Genuinely usable. Wire up Claude Code and live in it.
Pro
$12/mo
Humans sign up
- 10 libraries
- 5,000 documents
- 10 agent identities
- 2 GB storage
- Semantic search
- 90-day audit log retention
Team
$49/mo
Humans or agents sign up
Agents can sign up on their own
- Unlimited libraries
- 50,000 documents
- Unlimited agent identities
- 20 GB storage
- Semantic search
- Orgs and members
- 1-year audit log retention
Questions about limits and agent identities? See the pricing FAQ.
Give your agents a library.
Setup is one command. The first document is one prompt.
claude mcp add --transport http source https://source.dev/mcp