Less webpage. More usable context.¶
Live URL → compact Markdown for AI agents — or a typed
ok:false. Strips chrome, fits a token budget, keeps the source.
Does not invent page text from model memory.
Webpage noise
<header>…</header>
<nav class="site-nav">…</nav>
<script src="bundle.js"></script>
<aside class="sidebar">…</aside>
<footer>…</footer>Source-linked Markdown
1 # Web context without the chrome
2
3 Agent infrastructure · 8 minute read
4
5 AI agents rarely need the whole
6 interface of a webpage. They need
7 the useful text, its structure, and
8 enough source information to explain
9 where the material came from.
10
11 ## The hidden cost of a page readsource: representative-input.html
- This demo · HTML
- 5,297 B
- This demo · Markdown
- 1,736 B
One inspectable demo page — not an average, not a promise. Real sites vary (sometimes much more chrome, sometimes almost none). Bytes, not tokens.
curl -fsSL https://raw.githubusercontent.com/ContextForgeAI/occam/v1.3.0/scripts/get-ff-occam.sh | bash
occam connect): npx ff-occam@1.3.0.
What Occam is for (skills, not a tool pile)¶
Occam’s eighteen core tools combine into a few finished abilities. Default
profile reader exposes eleven day-to-day tools; authoring and the full
catalog need OCCAM_PROFILE=full.
| Skill | Situation → result | How tools join |
|---|---|---|
| Read under a budget | Need page material without burning context → compact Markdown or typed ok:false |
Optional search / map / probe → occam_transcode or one-call occam or occam_digest; budget via occam_client_capabilities; cheap re-read with if_none_match / diff_against |
| Learn a site recipe | Default extract fails on a site you will revisit → local playbook reused later | Agent loop: heal → lint → save (full profile); later resolve / transcode / occam_extract_knowledge. Notes guide the agent; they are not an automatic price converter |
| Bind evidence | Need citations or proof of what was returned → integrity / block match / heuristic attest / canary | Receipt + occam_verify, occam_claim_check, occam_attest, canary, occam_dataset_export — different questions, not one “truth” button |
| Size the tool surface (opt-in) | Match the tool menu to observed agent behaviour → narrower or fuller tools/list |
OCCAM_EXAM_MCP=1 + submit a behaviour record; the host grades the record, it does not run the exam for the agent |
Local-first defaults, private-URL blocks, and typed unknowns apply across skills. Opt-in volume/time tools (batch, watch, crosscheck, atlas, browser actions) stay behind env flags — see configuration.
Full knobs: Why Occam · agents: llms.txt ·
repo door: README.
Inspect one demo (not a benchmark)¶
The product claim is usable context or an honest unknown — not a fixed “% smaller” rate. Chromed marketing pages often shrink a lot; a clean docs page may barely shrink. We publish one stable fixture so you can open the input, run the same call, and check the method.
<header>…search…</header>
<nav>…repeated links…</nav>
<main>
<article>Web context without the chrome…</article>
<aside>…newsletter…related reading…</aside>
</main>
<section>…cookie notice…</section>
<footer>…company links…</footer>5,297 UTF-8 HTML bytes · this page only
# Web context without the chrome
AI agents rarely need the whole interface of a webpage.
They need the useful text, its structure, and enough
source information to explain where the material came from.1,736 UTF-8 Markdown bytes · this page only
For this demo page the Markdown body is shorter than the HTML body (5,297 → 1,736 UTF-8 bytes). That ratio is a property of the fixture, not a product KPI — do not cite it as “Occam saves 67%.” No tokenizer was used; not a token or quality claim.
Method and source revision
Runtime source revision `b3c212c6d9e193619b6e8663148bd53932a0acc0`. Full method, complete input/output, the minimal live smoke proof, controlled failure, and reproduction scripts: [current proof bundle](examples/current-proof/README.md) and three recorded jobs in [golden workflows](examples/golden-workflows/) [context packs](examples/context-packs/), and [site research](examples/site-research/).The cost grows with every source¶
An ordinary page can contain navigation, scripts, repeated interface text, cookie controls, footers, and large amounts of raw markup around the part an agent actually needs.
Passing all of that through wastes a local model's limited context and makes larger models work around irrelevant text. A search result can help find the source; Occam focuses on reading the chosen source and preparing usable page content for the agent.
The measured fixture above makes that transformation inspectable. The proof
bundle also retains example.com as the minimal live smoke case so the richer
marketing example does not replace the smallest deterministic contract check.
A useful result — or an explicit unknown¶
Readable content with the source URL and result metadata.
# Example Domain
This domain is for use in documentation examples…Page content is unknown — do not invent it from memory.
{
"ok": false,
"failure": {
"code": "private_url_blocked",
"message": "Private or local URLs are blocked."
}
}This predictable failure is a trust feature, not the main reason to try Occam: start with the successful read. The proof fixture includes the controlled private-destination case above.
Choose your workflow¶
Three routes. Pick one, then open the matching Quick Start path for the exact prompt.
Connect Cursor, Hermes, or another MCP-capable app, then type the first prompt in a new conversation.
Open AI app paths →Tool-capable local model via Occam’s terminal chat — Ollama is a model runtime, not an MCP host.
Start local Ollama chat →Wire your own MCP client with the generated launch snippet and API contract.
Developer integration →Choose a detailed path and get the exact first prompt · Host onboarding · Workflow gallery · See host validation tiers
How Occam works¶
Occam starts with a lightweight local read and can use a local browser when the page requires it. It then returns the useful content and metadata through one agent-facing contract. Focus, budgets, structured extraction, sessions, and verification are available when the task needs more control.
How Occam works · Read one page · Choose a tool
Why Occam¶
Not “another web fetch.” Occam is the local honesty-and-evidence layer between your agent and the public web: acquire real material (or refuse), fit a token budget without LLM summarization, optionally learn a site recipe, and optionally attach checkable integrity aids.
| Generic fetch / memory | Occam |
|---|---|
| Empty HTML or invented text | Live Markdown or typed ok:false |
| Burns the context window | Budget + focus prune (deterministic) |
| No proof of what was returned | Optional Receipt v1 → verify (integrity vs a key — not truth) |
| One opaque “read” | Skills: budgeted read, playbook authoring, evidence, opt-in exam surface |
Why Occam — full flashcard · Choose a tool · Ask AI / agent prompt
Trust and local control¶
Local-first is not an absolute “never cloud” claim. The origin website is a network source, and explicitly configured search, proxy, or remote transport providers can change the boundary.
Optional Receipt v1 artifacts can check returned-byte integrity against a supplied key. They do not prove truth, identity, or authentic origin.
Trust and security · Receipts · Installation safety · Reference overview
Deep hubs (after the product path): TRUST · CASCADE · SEARCH · MCP map · Capabilities · Feature inventory
Get your first Occam result¶
Install Occam, choose the path for your AI application or local runtime, and finish with a real page read.
Get your first result Inspect the proof
Explore (product path first)¶
Why Occam · Task router ·
Tools · Tools reference ·
MCP API · Configuration ·
Workflow gallery · Examples ·
Recipes · Operators ·
Troubleshooting · FAQ ·
Reference overview · Ask AI ·
llms.txt
Optional deep / engineering (not the onboarding funnel): Handbook · Experimental · Architecture decisions · Testing results · Install failure regressions · Proof-of-read protocol
Research drafts under docs/research/ remain on disk for maintainers but are off the
product funnel (not linked from the primary Explore path). docs/rc2/ stays MkDocs-excluded.
All top-level pages¶
Compact index (gate + scanning). Narrative map: Reference overview.
acquisition · ask-ai · CAPABILITIES · CASCADE · choosing-a-tool · concepts · configuration · datasets · documentation-map · experimental · failure-codes · faq · getting-started · how-occam-works · install · materialization · mcp-hosts · MCP_MAP · networking · operators · playbooks · quality-baseline · quick-start · receipt_verification · receipts · recipes · roadmap · SEARCH · sessions · tools-reference · transports · troubleshooting · TRUST · trust-and-safety · what-is-occam · why-occam