One link, one Library
Each Connect link reaches exactly one Library. Tools on offer depend on the caller's role there: Reader or Publisher.
Your agents answer from
what you published.
Help desk runbooks, security policies and incident procedures, with citations and review-by dates. Agents can draft changes, but nothing goes live until you approve it.
Demo data is fictional.
Works with
CursorClaude CodeClaude Desktopclaude.ai & hosted web connectorsLibreChat (Reader key)
Agents draft. People publish. Readers see only what's approved.
MCPEncode stores the facts your agents rely on, like policies, procedures, runbooks and course material, as Items in Libraries. Agents connect to one Library to look Items up and suggest edits. With Review Lock on, those edits wait in Pending until a person publishes them.
- 01
Add a Library
Pick a starter (Policy, Procedure, Documents) or Custom. A person approves the schema once.
- 02
Connect an agent
Paste the Library's Connect link into your MCP client. Sign in, click Allow. That connection sees this one Library only.
- 03
Agent drafts
create_item/update_itemland in Pending under Review Lock. The published version stays live. - 04
A person Publishes
A Publisher reviews the diff and Publishes, or Cancels. Readers and agents get the new version only then.
Try it: Review Lock vs. Review Unlock
Flip the write mode, then have the agent save an edit.
Review Lock (default)
Waiting for the agent…
Live — what Readers & agents see
Visitor Wi-Fi policy · v1
Pending — Publishers only
—
Built so agents can't quietly change the facts.
OAuth 2.1 built in
Clients register themselves, people sign in and click Allow. Or plug in an external OIDC issuer such as Zitadel. Disabling an account takes effect on the next request.
Citations with version
Every answer carries the published version, date, publisher and a link, ready to quote as cite_as. Pending drafts are never cited.
Audit log
list_audit shows who called which tool in a Library, for Publishers and the Site admin.
Review-by freshness
Runbooks and policies get a review-by date. Stale answers carry a warning; strict mode refuses them and names the owner. A Review queue has one-click Still accurate.
Agent-suggested aliases
"We got phished" finds the incident runbook. Agents suggest labels and aliases with suggest_labels; a person approves them like any edit.
The MCP tools
| whoami | Identity, library, role |
| list_items | Look up by question or browse; one clear match returns the full published Item |
| open_item | Full Item by id (published version; Publishers may read the draft) |
| search_shelf | Published summaries across the Library's Shelf |
| create_item · update_item | Save to Pending (Review Lock) or publish at once (Review Unlock) |
| suggest_labels | Suggest labels and aliases for a published Item; a person approves |
| delete_item | Remove an Item |
| create_library · propose_table · propose_field | Design a Library; a person approves the schema |
| list_audit | Who called which tool in this Library |
Anywhere an agent needs the approved answer.
Help desk runbooks
VPN, lockouts, printers, laptop setup. Aliases and typo correction match how users actually ask; a Review queue keeps runbooks current.
Security policies
Policies with owner, version, effective and review dates. Agents cite the version in force, and strict mode stops them quoting one past review.
Incident response
The first-hour runbook your agent hands out is the one a person approved, not a draft. Edits wait in Pending.
Beyond IT
The same model works for support playbooks, HR handbooks, product docs and course material. Entity keys let repeated syncs update one Item.
Self-host in four steps
Needs Docker with Compose, and openssl.
1. Clone
git clone https://github.com/hkrlogic/MCPEncode.git && cd MCPEncode2. Create .env with fresh secrets (never overwrites a value)
scripts/init-env.sh3. Start the app and Postgres
docker compose up -d --build4. Open http://127.0.0.1:3000/
Create the first account with
SETUP_TOKENfrom.env. It becomes the Site admin. Add a Library, add an Item, Publish.
Then connect your agent
Open a Library → Connect → copy the Connect link.
claude mcp add --transport http my-library <Connect link>
Settings → Tools & MCP → add a server
with the Connect link as url
with the Connect link as url
npx mcp-remote <Connect link>
# as the server command
# as the server command
Connect from claude.ai or another web connector;
a Site admin approves the app's host once,
or set OAUTH_REDIRECT_HOSTS=claude.ai
a Site admin approves the app's host once,
or set OAUTH_REDIRECT_HOSTS=claude.ai
Your client opens a browser to sign in and Allow. Use 127.0.0.1, not localhost.
Questions
Can an agent publish or approve its own changes?
No. MCP has no publish, approve, write-mode or revert tool. Under Review Lock, a person Publishes in the web UI.
Do Readers see an Item while it's being edited?
They see the last published version until the edit is Published. Draft text is never returned to them.
What happens when a runbook goes stale?
Past its review-by date, answers carry a warning naming the owner. Turn on strict mode and agents refuse it instead. The owner clicks Still accurate in the Review queue, or edits and Publishes.
Is it hosted, or do I run it?
You run it: one Node.js 22 app plus PostgreSQL 16+, via Docker Compose.
How do agents authenticate?
Built-in OAuth 2.1 sign-in, or an external OIDC issuer such as Zitadel. A Reader key gives a shared agent read-only access without sign-in.
What if a Connect link leaks?
Rotate it under Connect. The old link stops working; Access grants stay.