METANTEL · INTENT-COHERENT AGENT MEMORY

Intent-coherent, agent-paced memory for AI teams.

Human-like memory and HPC-inspired collaboration. Give the coding agents you already pay for—like Claude Code—the right context handed over at agent-controlled volume, fewer tokens burned, and conflict-free team awareness.

Exploration’s reach at retrieval’s cost.

Join Early Access · 45% Off → See What You Gain ↓ Read the Benchmarks →
Connect Harnesses, not just MCPs Runs on the Claude Code subscription you already pay for—zero new inference bill.
Flow Console · cal.com — 2 agents active · EDIP Feeder connected
MOESI Coherent
EDIP Feeder · Step 1 3 of 19 judged documents served · 4.6 s
✓ src/client.ts
86% match
✓ src/config.ts
79% match
✓ session:a191da#turn8
92% memory
Agent-paced in rounds: session remembers what it served; agent pulls until it has enough.
HPC Cache Coherence · Live Branch State 0 Collisions
Laptop A · Agent 1 src/client.ts → get() Editing: Adding retry backoff
Laptop B · Agent 2 src/client.ts → get() Sees Agent 1 intent & uncommitted diff before editing
−47%
Long-session coding cost

40 coding requests on Sonnet 5.5 ($9.78 → $5.13). Compounds from −19% in half 1 to −63% in half 2. (−74% on Haiku 5.5).

4th of 26
Public EnterpriseRAG-Bench

500 questions over 511,962 company records. Ahead of OpenAI, Amazon, Azure & Google search; #1 on conditional questions.

0 Collisions
Function-level MOESI coherence

Overlaps surface before the edit, not at git merge. 0 failed merges across 132,920 simulated 8-agent merges.

$0 AI Bill
Bring the agent you already pay for

Runs on your existing Claude Code subscription and licence. We sell the harness and memory—never markup on tokens.

VALUE ADDITION · WHO YOU ARE & WHAT YOU GAIN

What changes for you on day one.

Whether you are a solo builder trying to stop 300K-token context bloat, an engineering lead coordinating parallel agents across laptops, or an enterprise governing company-wide intelligence—here is the exact Pain Point, Solution, and Why Metantel wins.

For Individual Developers

Stop paying your coding agent to re-read 300K tokens on every turn.

Personal · $10 / mo
The Pain Point

Exploratory agents burn tokens re-reading the same files

Tools like Claude Code and Codex search, grep, and follow file names from scratch on every prompt. In a long coding session, peak context balloons to 313K tokens and gets re-read on every call—driving up API costs and slowing down responses.

The Solution

Agent-paced memory that feeds only what the task needs

Flow Console + EDIP connects to the Claude Code subscription you already pay for. Instead of blind grep loops, EDIP serves pre-linked, Librarian-judged evidence in rounds until your agent has enough—and lets you graduate strong sessions into persistent AI Workers.

Why Metantel?
−47% Cumulative session cost on Sonnet 5.5 (40 requests)
  • −19% on requests 1–20, compounding to −63% on requests 21–40 ($9.78 → $5.13)
  • −74% across 39 coding requests on Haiku 5.5; 8 of 8 short 5-request sessions cheaper (−16% net)
  • Peak context cut from 313K to 117K tokens—with zero new AI inference bill from us
For Engineering Teams

More agents, not more merge conflicts and coordination meetings.

Team · $25 / user / mo
The Pain Point

Parallel agents work blind and collide at git merge

Every engineer uses an AI assistant now, but each agent only sees its own laptop and window. Studies show 27.7% of agent-authored PRs hit merge conflicts because agents duplicate work, overwrite functions, or break assumptions another teammate’s agent relies on.

The Solution

HPC MOESI cache coherence brought to AI coding agents

A lock says stop; coherence says keep working, and intent settles the conflict. In Flow Console, every agent knows who else is working, on what and why. When two agents reach the same function on the same branch, one sees the other’s uncommitted change and intent before editing.

Why Metantel?
0 / 132,920 Failed merges across 8 coordinated agents in simulation
  • Overlaps surface before the edit, not in PR review—humans step in only when agents disagree
  • Direct cross-machine agent-to-agent messaging with MOESI function-level locks as the safety net
  • One shared memory across devices: what one developer’s agent learns, every teammate’s agent can use
For Enterprise & Orgs

Multi-hop precision across code, tickets, and docs—scoped to your org chart.

Custom & Self-Hosted
The Pain Point

One-shot RAG and raw MCPs dump unjudged noise and lack governance

Enterprise search, standard RAG, and raw MCP tool calls run one query, dump a fixed bundle of noisy snippets, and stop—cutting off evidence past one hop and burning your model’s context. Meanwhile, legacy internal web apps have no governed harness for agents.

The Solution

Connect Harnesses, not just passive MCPs: EDIP + Workflows + Agentify

Instead of exposing dumb one-shot tools, Metantel connects intelligent Harnesses. EDIP builds multi-hop links ahead of time (Connectome) and uses the Librarian Retrieval Harness to judge every step in rounds. Wire data into plain-English decision endpoints with EDIP Workflows, and turn existing web apps into governed agent APIs with Agentify.

Why Metantel?
#1 On conditional enterprise questions · 4th of 26 overall (EnterpriseRAG-Bench)
  • Outscored OpenAI, Amazon, Azure, and Google search over 511,962 company records (500 benchmark questions)
  • Strict org-ladder scoping: 302 of 302 authorized facts delivered, 0 of 573 leaked across boundaries
  • No repo checkout required on disk (85 vs 88 score); git pushes become searchable in ~25 seconds
FOR A CUSTOMER TEAM · DAY ONE IMPACT

Side-by-side: Today vs. With Metantel

Connect Harnesses, not just passive MCPs
When Today (Without Metantel) With Metantel (EDIP + Flow Console) Measured Evidence
An agent needs the "why" Searches, guesses, or interrupts a senior engineer Pulls the memo, ticket, and thread, with verified sources #1 on conditional questions, public leaderboard
A long coding session (40 requests) Whole conversation re-read on every call (313K peak context) Context kept small (117K peak); EDIP feeds only what the agent needs −19% → −63% as session runs; −47% overall on Sonnet 5.5
Two agents, same code, two laptops Discovered at git merge; fixed manually by people or separate agents Told before editing; held only on the exact same function via MOESI coherence 0 failed merges across 132,920 simulated 8-agent runs
THE MEMORY PERFORMANCE VS. PRECISION PROBLEM

Two ways to feed an agent. Both break.

A pipeline is only as good as its bottleneck: finding evidence across hops, and delivering it to the model whole.

Exploratory agents

Claude Code, Codex

Question
Search, read, follow names
Repeat: high context re-read

Reach, at high cost. Company docs have no file names to follow, and context bloats to 300K+ tokens.

Metantel: the hybrid

EDIP and our harnesses

Question
Connectome: links built ahead
Librarian: every step judged
Rounds until agent has enough

Exploration’s reach at retrieval’s cost. What agents use flows back into shared team memory.

Precision retrieval

Glean, RAG, memory APIs

Question
One search
One fixed bundle, then stop

Cheap, but blind past one hop. The pipeline picks the context size, not the agent.

EDIP · THE MEMORY LAYER

Multi-Layered Memory Distribution.

Live short-term
Every agent’s intent and uncommitted edits, per repo and branch, held in live memory.
Sessions experience
Every session’s context and interactions stored in EDIP to keep serving important instructions. Sessions graduate into AI Workers.
Knowledge long-term
Code, docs, and notes. The Connectome links them automatically; the Librarian Retrieval Harness judges every step.
Team organisational
Facts written as per the org ladder. Datafeed strictly scoped as per team-level access.
CONNECT HARNESSES, NOT JUST MCPS

Why agent-paced harnesses beat one-shot tools

One-shot: raw MCPs & search results (Glean, Augment, RAG)

A raw MCP or RAG tool returns a fixed bundle and forgets it. The coding agent pays tokens to sort through the noise.

One-shot: a finished answer (Unblocked)

A pipeline summarizes prematurely. Crucial multi-hop evidence is cut before the strong reasoning model ever sees it.

Agent-paced Harness: EDIP, in rounds

Connect a stateful Librarian Harness, not a dumb pipe. Serves judged whole documents across rounds; remembers what it served and tells the agent what’s left.

PRODUCT LINE · LAND, TEST, EXPAND

One memory. Three workloads.

Everything starts with EDIP—our org-scoped, multi-modal, coherent retrieval and memory pipeline. Above EDIP sit three focused harnesses that take your team from multi-agent coding, to automated browser QA, to governed enterprise agent APIs.

PRIMARY PRODUCT · BUILD Beta

Flow Console

We brought HPC cache coherence to AI agents.

Where coordination bites. Run multiple coding agents across teammates and machines on the AI provider you already pay for—like Claude Code—inside an active orchestration harness with shared agent-paced memory and zero merge collisions.

Pain: Agents working alone burn tokens and break each other’s code
Solution: An active orchestration harness: MOESI cache coherence + EDIP memory
Why Us: Compounding token savings and conflict-free shipping on your existing AI bill
  • Intent settles conflicts before the edit; MOESI function-level locks underneath as the safety net
  • −19% cost on requests 1–20, compounding to −63% on requests 21–40 (−47% overall on Sonnet 5.5)
  • One memory across devices: what one developer’s agent learns, every teammate’s agent can use
  • Runs on the Claude Code subscription you already pay for—no new AI inference bill
Explore Flow Console →
EXPERIMENTAL · QA HARNESS Experimental · Lite on npm

Artel QA

Reads the source, drives a real browser, rules from evidence.

Our experimental QA harness that maps every interactive control directly from your source code, exercises your running application in a real browser, and adjudicates pass/fail verdicts strictly from network, DOM, and database evidence. Artel QA Lite runs locally on your own agent via npm.

Pain: AI-generated code ships faster than humans can test, and raw browser MCPs guess
Solution: A white-box QA Harness: source inventory + deterministic evidence adjudication
Why Us: Zero-config verification that feeds directly back into Flow Console and matures into Agentify
  • Reads source code first so no untested control or route can hide
  • The Harness adjudicates from hard evidence: a "Saved" toast with no network write is marked Refuted
  • Caught 5 of 5 broken controls and accused 0 of 4 healthy ones ($0.36 run cost)
  • Artel QA Lite runs on your own agent via npm; full edition integrates with EDIP memory
Explore Artel QA →
SECURE WEB-APP HOSTED MCPS Under Development

Agentify

The Web Is the API.

Matured from the Artel QA Harness. Turn the internal and customer-facing web applications you already run—ERPs, CRMs, helpdesks, and homegrown admin panels—into scoped, audited agent APIs and hosted MCPs that AI agents can call safely.

Pain: Hand-rolled MCP wrappers take months to build and lack enterprise guardrails
Solution: Governed, web-app hosted agent APIs matured directly from the Artel QA Harness
Why Us: Land with coding & QA harnesses, expand to enterprise-wide operational agents
  • No integration project or backend rewrite: your existing web app becomes the governed interface
  • Matured from Artel QA: builds directly on verified control and network maps
  • Strictly scoped: each agent receives only the specific actions its role permits
  • Full audit trail: every invocation records who asked, what ran, and what changed
Explore Agentify →
FOUNDATION LAYER · EDIP & EDIP WORKFLOWS (BETA) Beta · Open to Partners

EDIP: The Retrieval, Memory & Plain-English Decision Pipeline

The EDIP memory engine that feeds coding agents in Flow Console also powers EDIP Workflows—letting teams wire any multi-modal data (code, notes, reports, tickets, images) into plain-English decision boards with confidence scores, evidence citations, and one-click REST/Python SDK endpoints.

1. Wire parts on a board Data in, decisions, shaping, results and alerts out—or describe the board in a single sentence.
2. Every decision is a question Asked of every item, returning a calibrated confidence score and the exact evidence behind each answer.
3. Publish as an endpoint Any internal app or agent can call a workflow through the REST API or Python SDK.
Explore EDIP & Workflows → Propose an Enterprise Pilot →
COMPANION INTERFACES

Control on the web. Dispatch from your phone.

Artel Workspace Live Beta

Manage your organization’s access ladder, data connectors, EDIP Workflows, docs, and tasks in one place—where humans and coding agents read and write to the same grounded project state.

View Artel Workspace →
Elenchus In Development

The mobile command node for the Metantel platform. Dispatch Flow Console tasks remotely, query EDIP with full evidence citations on the go, and approve agent decisions when a human is needed.

View Elenchus →
EVIDENCE · REAL CODING SESSIONS & PUBLIC BENCHMARKS

Long coding sessions: the saving compounds.

Measured side-by-side on the exact same agent, model (Claude Sonnet 5.5 & Haiku 5.5), machine, repository, and 40 coding requests (features, fixes, refactors, tests, and docs)—with EDIP’s own cost included.

Cumulative Claude Code Session Cost (40 Requests · Sonnet 5.5)
Claude Code alone ($9.78) With Flow Console + EDIP ($5.13)
$4.65 saved per session (−47%)
$10 $8 $6 $4 $2 $0 Req 1 10 20 30 40 Requests 1–20: −19% Requests 21–40: −63% $9.78 $5.13 $4.65 saved

Same agent, model, machine, repo and requests, side by side (Oct 2026); cost as Claude Code reports it, EDIP’s cost included. Acceptance checks: 28/32 vs 32/32 on Sonnet (3 of 4 misses: the agent asked before a design call), 30/31 each on Haiku.

−47%
40 coding requests on Sonnet 5.5

Features, fixes, refactors, tests, and docs. Peak context reduced from 313K to 117K tokens re-read on every call.

−74%
39 coding requests on Haiku 5.5

Plus 8 of 8 short 5-request coding sessions were net cheaper (−16% net).

−51% net
40 Q&A questions, no edits ($0.32 → $0.16)

37 of 40 cheaper with identical fact accuracy (93.1% vs 91.9%). Even with no repository checkout on disk, Claude Code answered through EDIP nearly as well (85 vs 88).

COMPETITION · WHY US?

Each rival owns one piece. Metantel unifies all three.

One-shot tools return a fixed bundle and forget it. No competitor we found exposes both an agent-paced continuous feeder and distributed function-level agent collaboration.

Company Category / Backing Context Reaches the Agent As Agent-Paced Feed Coordinates Before Merge Live Updates Mid-Task
Glean $7.2B valuation Search results No, one-shot No No
Unblocked $30M raised A finished answer No, one-shot No No
Augment $252M raised Search results No, one-shot No No
SageOx $15M seed Notes at session start No, one-shot Advisory, file level No
Claude Code Projects Anthropic Its own search No, one-shot No, git merge One machine
GitHub Agent HQ Microsoft Its own search No, one-shot No, at the PR No
Mem0, Zep Memory APIs Stored memories No, one-shot No No
Metantel Pre-seed Judged evidence, in rounds Yes Yes, function level Yes, across machines

Sources: company announcements and public documentation, 2024–2026. Benchmarks scored against the 28 Aug 2026 EnterpriseRAG-Bench leaderboard and internal 40-request / 40-Q&A test suites (Oct 2026).

BUSINESS MODEL · NO INFERENCE BILL

Bring the agent you already pay for.

AI workers run on your own Claude Code subscription and licence. Start with a Free trial of our harnesses, step up to Personal ($10/mo) for solo builders, or land a squad on Team ($25/user/mo) for shared memory, cross-machine intent collaboration, and the org ladder.

Compare Full Plans & Features →
Free $0 Trial versions of our harnesses
Personal $10 / mo Flow Console, Artel QA & EDIP memory
Team $25 / user Shared memory, intent coherence & org ladder
EARLY ADOPTERS · LIMITED SEATS

Be first in. Lock in 45% off for two years.

Join the early access waitlist for Flow Console, Artel QA, and EDIP memory. Early adopters keep a 45% discount on the baseline price of every Metantel subscription for their first 24 months.

  • 200–500 early adopter seats: The 45% discount closes once the early adopter allocation is filled.
  • Founding AI builders cohort: We are inviting a select cohort of active AI builders to use Metantel free. Tick the box in the form and share your work to be considered.

Limited seats. One per person, verified when you create your account. Duplicate, alias and disposable addresses aren't eligible. Fair-use limits apply.