What is Occam?¶
Your AI is good at reasoning. It is bad at silently inventing what a web page said.
Occam is a locally run host that sits between your AI and the public web. It acquires real page content, shapes it for a model’s context window, returns a typed refusal when the content is unknown, and can optionally attach integrity artifacts you can verify later against a key.
Your AI
|
| MCP / CLI
v
Occam (local host)
|
+-- acquires the page (HTTP → browser → optional managed)
+-- materializes token-budgeted Markdown / structured fields
+-- returns ok:false when content is UNKNOWN
+-- optionally signs what it produced (integrity vs a key)
The problem¶
Without a dedicated extract layer, agents either:
- skip the fetch and hallucinate from training memory, or
- dump raw HTML into the context window and waste tokens
Neither is trustworthy for research, citations, or automation.
What happens on a normal read¶
- The AI asks Occam to read a URL (usually
occam_transcodewith onlyurl). - Occam acquires the page now through a gated ladder (HTTP first; browser when the HTTP result is unusable; optional managed provider only after both local attempts fail).
- Usable content is materialized: budgeted, optionally focused, optionally structured.
- On success:
ok: trueplus Markdown (and optional structured sidecars). A signed receipt may be attached. - On failure:
ok: falsewith a typedfailure.code— content is unknown, not guessed.
Live extraction is the default. An opt-in response cache (cache_ttl_s) can replay a prior materialization on the same machine — it is not a CDN, and it is off unless you set it.
Beyond one URL¶
Occam also supports:
- Discovery before a full read — probe, map, search
- Several URLs in one call — digest
- Site recipes (playbooks) — resolve, heal, lint, save
- Typed fields — extract_knowledge (needs a schema)
- Evidence lookup / citation assessment — claim_check, attest (heuristic — not truth proof)
- Integrity checks — receipts, verify, dataset export
- Experimental opt-ins — watch, crosscheck, batch, failure atlas
What it does not promise¶
| Not this | Reality |
|---|---|
| Proves the page was true | Signatures prove integrity relative to a key |
| Proves origin authenticity / identity | Local self-minted key; no PKI / registry |
| Trusted timestamps | Signer clock only (optional TSA is separate and limited) |
| CAPTCHA solving | Detects walls; does not bypass them |
| Universal npm install | npm is not a GA 1.0 channel (experimental RC packages ff-occam / @ff-occam/mcp exist) |
| Cosign as page-content truth | Published v1.0.0-rc.3+ use signaturePolicy=required-cosign-v1 (SHA-256 always + Cosign fail-closed). Cosign proves release authenticity vs the workflow identity — not page truth. v1.0.0-rc.2 remains SHA-256-only |
| Marketplace “trusted auto-merge” | Community automation is operational machinery, not a trust guarantee |
Who it is for¶
| You | Occam helps you |
|---|---|
| Person using Cursor / Claude / Codex / Hermes | One install, then ask the agent to read the web |
| Operator | Local install, connect, doctor, sessions, packaging |
| LLM agent | Typed tools, honest failures, compact llms.txt map |
| Auditor | Verify receipts offline without trusting the chat transcript |
Next¶
- Why Occam — advantages and knobs — read this first (human + agent)
- Quick Start
- How Occam works
- Trust & Safety
- Handbook