› how long do we retry failed webhooks?
Retries run for 3 days with linear backoff.
No source. Sounds right. Is wrong.
Build a knowledge base
your agents can use.
macOS · Linux
From your docs
to a cited answer.
Point setup at the Markdown you already keep. Then ask a question. Search returns the exact file, the line range, and a content hash your agent can cite.
Agents guess.
Give them receipts.
Without your project's knowledge, an agent fills the gap with something plausible. With Open Knowledge, it reads the same Markdown your team maintains, and every answer points to the lines it came from.
› how long do we retry failed webhooks?
24 hours, with exponential backoff.
↳ imported/delivery.md:6-11
Four commands.
One knowledge base.
No account, no server, no model required. Everything runs locally against plain Markdown in your repository.
$ okn setup Wiki --from docs --mode copy Setup plan Mode: copy Documents: 3 Changes: 9 create 3 add minimal frontmatter 0 delete Knowledge base created at ./Wiki. Documents 3 Validation READY Search index READY (2 hits)
$ okn search Wiki "how long do we retry?" # Open Knowledge Context Status: managed Context: 1410 / 2400 estimated tokens Sources: 12 ## 1. Why we retry webhooks Source: imported/delivery.md:6-11 Locator: okf+sha256://3b09…/imported%2Fdelivery.md#78fa… Relation: direct Retry failed deliveries for 24 hours with exponential backoff.
$ okn view Wiki Open Knowledge view: http://127.0.0.1:4185/wiki/ root /your-project/Wiki Press Ctrl+C to stop. # In the browser, a WebMCP agent can: # openknowledge_search find pages # openknowledge_apply_edit change the shared draft # openknowledge_save_document save after you review
$ okn check Wiki Knowledge status Structure READY 6 Markdown files validated. Links READY Local Markdown links resolve. Freshness NOT CONFIGURED No audit baseline is configured. Retrieval NOT CONFIGURED No eval dataset is configured. Publication READY Content checks passed. Overall READY
Start local.
Grow into production.
The same Markdown bundle adds checks, CI, and a production runtime when other people and agents depend on it. No format migration on the way.
Answers with receipts
Search packs the best sections into a token budget. Each one carries a content-addressed locator, so an agent can cite exactly what it read.
Source: imported/delivery.md:6-11 Locator: okf+sha256://3b093f35…/delivery.md#78fa9eec… Context: 412 / 2400 estimated tokens
Every agent, one wiki
A read-only MCP server for any client. Setup can hand the job to the agent you already use.
"project-knowledge": { "command": "okn", "args": ["mcp", "/path/to/Wiki"] }
- Claude Code
- Codex
- OpenCode
- any MCP client
Agent proposes. You decide.
Agents edit a shared draft in the local viewer. You see the diff before anything is saved. History stays in Git.
@@ -10,0 +11,2 @@ A receiver can be unavailable during a deploy. + +Honor the Retry-After header on HTTP 429.
Knowledge CI
Pin evidence, watch sources for drift, and test retrieval against known questions on every pull request. Claim changes wait for a person.
✓ Open Knowledge checks passed ✓ eval · regressions gate 0 regressed ✓ evidence pinned src/relay/retry.ts ● claim change needs human approval
Ship knowledge like code
Publish a static viewer, llms.txt, and a connect manifest. The production MCP runtime serves immutable generations with explicit health, and rolls back in one command.
gen 12
gen 13
gen 14active
gen 15degraded
↶ gen 15 degraded, so the runtime rolled back to gen 14
Nothing to lock you in
Plain Markdown with YAML frontmatter, in your Git repo, in the open OKF format. Remove the tool and your knowledge still reads fine.
One lifecycle.
No migrations.
Ordinary docs need only Markdown, checks, and search. Add Knowledge CI and a production runtime when agents in production depend on the answers.
- sourcemarkdown + git
- checkdeterministic checks
- reviewevidence-backed PR
- publishproduction MCP
- looprollback · usage · next change