TwinDragonRegister interest

02:14 · usage limit reached

Your agent hit its limit at 2 a.m. The next session starts from zero.

More features from the same subscription.

Fewer retries, less money on tokens, and work that survives the usage limit and the model switch.

Register interestBook 20 minutes with the founder

TwinDragon runs on every developer's machine, captures why the code changed (agent sessions, pull requests, tickets, chat; meetings later), and gives Claude Code and Codex only the context each task needs, with a source behind every line.

two sessions, one night
02:14  agent  usage limit reached. session ended.
02:14  state  41 turns · 820K input tokens re-sent
02:15  you    "it had almost finished the migration"

07:02  agent  new session. context: none.
07:02  agent  "what are we working on?"
07:03  you    re-explain the ticket, the PR, and
              the thing tried on Tuesday. reverted.

with TwinDragon
07:02  twin   context pack · 1,140 tokens · 6 sources
07:02  agent  resumes at turn 41. knows about Tuesday.

Illustrative. Not a recording.

The four costs

Agents are cheap per token and expensive per feature.

Four places the money goes when the agent does not know why the code is the way it is. Each number has a source; none of them is ours.

01

Token waste

Every turn re-sends the whole conversation. Most of what you pay for, the model has already read.

25 : 1
A typical 50-turn agent session re-sends about 1M input tokens against 40K output. — Vantage, April 2026
~62%
of the bill is re-sent context. — 30-team audit, indicative
02

Wrong implementations

An agent that starts without the team's decisions picks a plausible approach, not the agreed one, and then defends it.

39% lower
scores in multi-turn work when models take a wrong turn and never recover. — Microsoft Research / Salesforce, 2025
03

Repeated mistakes

The agent that rebuilt the reverted approach yesterday is back today with no memory of it, on every developer's machine.

66% · 73%
of developers use coding agents; 73% use them daily. — Stack Overflow Developer Survey 2026
04

Broken handoffs

One subscription, one usage limit, one model. When the limit hits or the team switches models, the next session starts from zero.

75%
of startup developers use Claude Code. — Pragmatic Engineer, February 2026

What the agent receives

Only the context this task needs. A source behind every line.

Instead of the whole history, Claude Code or Codex gets a short pack through MCP: what it must respect, what was tried and failed, where sources disagree, and what nobody has decided yet. Each line points at the pull request, session, ticket or message it came from, so the agent can open it and you can check it.

  • MUST — decisions in force.
  • PITFALLS — tried and reverted. Negative knowledge, the part no README keeps.
  • CONFLICT — sources disagree; the pack says which one wins and why.
  • UNKNOWN — no source. The agent asks instead of guessing.
context pack · repo: api · 1,140 tokens · 6 sourcesillustrative
task: "Add retry to the payment webhook handler"
MUST
– Webhook handlers are idempotent on `event_id`; the dedupe table is `webhook_events`. [PR #412 · S-0931]
– Retries use the shared `backoff()` helper, never an inline loop. [ADR-7 · S-0877]
PITFALLS
– Exponential backoff on this handler was tried and reverted: it hid a signature-verification bug for three days. [PR #388 → revert PR #391 · S-0802]
– `payment_intents` is partitioned by month; cross-month queries must go through the view. [PAY-214]
CONFLICT
– Ticket says retry up to 5×. Team chat after the incident agreed on 3×. Latest decision wins: 3×. [PAY-301 vs #payments 2026-09-14 · S-0944]
UNKNOWN
– No source says whether retries should emit a metric. Ask before adding one. [—]
sources: PR #388 · PR #391 · PR #412 · ADR-7 · PAY-214 · PAY-301 · S-0802 · S-0877 · S-0931 · S-0944 · #payments 2026-09-14

Illustrative example. Invented repository and source IDs; not from any customer.

How it works

Four steps, no change to how your team prompts.

  1. 1

    Install on every developer's machine

    TwinDragon runs locally next to Claude Code and Codex. Nothing to change in how people already work with their agent.

  2. 2

    Capture why the code changed

    Agent sessions, pull requests, tickets and chat are recorded as they happen. Meetings come later. The why is captured from every engineer at once, not reconstructed afterwards.

  3. 3

    Build a cited decision timeline

    Decisions, reversals and open questions are linked to the exact source. What was tried on Tuesday and reverted is kept, with the reason.

  4. 4

    Serve each task its context pack

    Through MCP, the agent receives MUST, PITFALLS, CONFLICT and UNKNOWN for the task in front of it, with source IDs. Not the whole history.

Pricing

Public prices. Nothing to request a quote for.

Per seat, per month, with the annual price under each. Not for sale yet: registering interest puts you in line for the first pilots.

Free

$0one seat
  • One seat
  • Session handoff
  • 1 repository
Register interest

Team

$19seat / month
$15 on annual
  • 2–50 seats
  • Shared team memory
  • Context packs for Claude Code and Codex
Register interest

Business

$39seat / month
$32 on annual
  • Everything in Team
  • SSO
  • Production-holdout proof engine
Book 20 minutes

Enterprise

from $49seat / month
100-seat minimum
  • Everything in Business
  • VPC deployment: $25K / year, only if you need it
Book 20 minutes
Savings guarantee

Savings guarantee on annual Business and Enterprise plans: if the 90-day holdout shows savings below 2× our fee, the next quarter is free.

Startup program

Startup program: 50% off for 12 months for companies under 3 years old with under $10M raised.

Register interest

Nine questions. Two minutes. Then a slot with the founder if you want one.

We use the answers to decide which teams to run the first pilots with, and nothing else. If you tick the interview, the next page offers a 20-minute slot straight away.

Where we are, plainly
  • Founder survey, August 2026: 1,000+ professionals from 376 unique companies; 202 companies asked to evaluate TwinDragon with their team (work email and agreed to a follow-up).
  • MVP about 50% built. The first end-to-end workflow works: ingest, cited decision timeline, context delivered via MCP.
  • No product users and no revenue yet. No benchmark results yet, so we publish no savings figure of our own.
No personal addresses
Company size, in developers
Agents you usePick all that apply
How you pay for agents
How often you hit a usage limit or lose context between sessions
What you would try first
Would you join a 20-minute interview?
Would you join a paid pilot?

We confirm by email (double opt-in). Nothing is shared with anyone else, and there is nothing to buy yet.

A note from the founder

Why I am building this

I am a Senior Staff Engineer at Thmanyah and ex-Amazon. On the side I am CTO of a six-person team where agents write most of the code. That second job is where TwinDragon comes from.

With six people and several agents, the expensive part is not the tokens. It is the agent that confidently rebuilds the approach we reverted last week, the session that dies at the usage limit with the migration half done, and the hour the next morning spent re-explaining what the ticket, the pull request and Tuesday's chat already said. Every engineer on the team pays that tax, separately, every day.

The idea is simple: capture the why while it happens, on every machine, and hand each task only the context it needs, with a source behind every line. The handoff between sessions and models is the first feature because it is the pain I feel at 2 a.m.

Where we are, plainly: a founder survey in August 2026 with 1,000+ professionals from 376 companies, 202 of which asked to evaluate with their team; an MVP about 50% built with the first end-to-end workflow working; no product users and no revenue yet. We publish no savings number until the benchmark exists.

If any of this sounds like your week, I would like twenty minutes of your time. I will ask questions, not pitch.

Fakhreddine Belgaied
Founder, TwinDragon

Book 20 minutesRegister interest