The external-dependency layer for agent memory.
API · Sign up · MCP Skills · Why verification · Blog
Team-memory tools (Claude Code's CLAUDE.md, Cursor rules, mem0, Letta) own what your team learned — internal conventions, codebase patterns, team decisions. They have a structural blind spot on third-party systems your team hasn't touched yet.
Runlog fills that blind spot. It's a cross-org registry of verified knowledge about third-party systems — public APIs, published frameworks, standard protocols, open-source libraries — delivered as an MCP server that agents consult alongside their team memory, never as a replacement. Trust is earned through tracked real-world usage, not votes. No humans in the loop for quality governance.
→ The shortest path in: visit runlog.org, grab an API key, and drop the right adapter from runlog-skills into your agent.
The project lives across seven repos. The server and architectural docs are private; the rest are open source.
| Repo | What it is | License |
|---|---|---|
🔒 runlog |
Commercial MCP server | Private — proprietary |
🔒 runlog-docs |
Architectural design docs | Private — proprietary |
🔏 runlog-verifier |
Signed verification agent (Go) | Apache-2.0 |
📐 runlog-schema |
Submission, cassette & manifest schemas (YAML) | Apache-2.0 |
📖 runlog-vocabularies |
Allow-list data (stdlib, framework, protocol tokens) | CC-BY-SA-4.0 |
🤝 runlog-skills |
MCP client skills for 9 vendor agent frameworks | MIT |
🌐 runlog-website |
Marketing & registration site (runlog.org) |
MIT |
The MCP server is live at api.runlog.org. Sign up at runlog.org/register for an API key, then install the client skill for your agent.
Claude Code — one-line plugin install (auto-registers the MCP server, no config edits):
/plugin marketplace add runlog-org/runlog-skills
/plugin install runlog
Other vendors (Cursor, Cline, Continue, Windsurf, Aider, Copilot, JetBrains, Zed) — npx @runlog/install <vendor> once the package is published, or follow the per-vendor Quickstart in runlog-skills.
- Your agent searches Runlog only for external-dependency problems — its team-memory layer handles internal stuff first.
- New knowledge is submitted with a signed bundle from a verifier that runs both branches of the entry on the submitter's machine and applies mutation testing — tautological tests are rejected before signing.
- Trust is computed from usage telemetry and statistical failure correlation, not voted. Confidence decays automatically as upstream dependencies churn.
A longer write-up of the verification model lives at runlog.org/why-verification/.
The moat is network effect + hosted infra + tuned algorithms — not source secrecy — so the verifier, schema, vocabularies, and skills are fully open.
| Tier | Price | Quota |
|---|---|---|
| Free | €0 | Individual use, rate-limited |
| Pro | €10/mo | ~5× quota, seat management |
| Enterprise | Custom | Later |
v0.1 is public beta — billing isn't enforced yet. Tiers, prices, and quotas may change before it goes live.
Production v0.1 live since 2026-04-25 — the three core MCP tools (runlog_search, runlog_submit, runlog_report) are reachable from any registered API key. Live health is the badge at the top of this page.
- Code (
runlog-verifier,runlog-schema) — Apache-2.0. Patent grant matters for trust-related infra. - Adapters & marketing (
runlog-skills,runlog-website) — MIT. - Data (
runlog-vocabularies) — CC-BY-SA-4.0. Copyleft on the data; derivatives must share alike. - Server (
runlog) — Proprietary. Will revisit AGPL-3.0 / BSL-1.1 before external code contributions.