Twigg Review (2026): Pricing, Features & Honest Verdict

Twigg is a hosted API that stores your LLM conversations, fits them to each model’s context window and routes every turn to the provider you name. Verdict: a sharp idea with clear pricing, still young and US-hosted. Best for: developers building multi-model chat or agent apps. Price: provider cost plus 5%, prepaid, $10 in welcome credits. Rating: 7/10

What is Twigg?

We found Twigg on the Product Hunt board for 16 September 2026, ranked eighth, with the tagline “The context layer you never have to build.” Anyone who has shipped a chat feature knows the layer it means. You design a table for messages and tool calls. You write the code that decides what to drop when the transcript outgrows the window. You write a translator for each provider’s message format. Then you bolt on token accounting so you can bill for it. Twigg sells all four as one API.

The model is simple. You create a chat once and get an ID. After that you send only what just happened, a prompt or a tool result, plus the name of the model you want. Twigg assembles the context, sends it to Anthropic, OpenAI, Google, xAI, Fireworks or OpenRouter, streams the answer back and records what it sent and what it cost. It is run by Twigg AI Ltd, a company registered in England. This is its second Product Hunt launch under the name. The first, in October 2025, carried the tagline “Git for LLMs”, so the product has changed shape within a year.

What Are Twigg’s Key Features?

Stateful chats you can move between models

In Twigg, the model is a field on each request, not a property of the conversation. A chat started on a Claude model can continue on a GPT or Gemini model by changing one string, and Twigg handles the translation between formats, including the reasoning artifacts each provider expects to see replayed. The live catalogue listed 42 models across six providers when we read it. The docs make the argument well: a provider’s own stateful API “stores your conversation inside the thing that generated it”, which is why you cannot leave. Twigg stores it beside the models instead.

Context budgets and automatic compaction

A context budget sets how much history a turn may carry. The formula is published: the context window minus reserved output, times one minus a headroom fraction that defaults to 10%. When history no longer fits, Twigg either summarises the older prefix or drops the oldest messages, your choice per namespace. By default it starts preparing summaries once a chat reaches 80% fill. Nothing is deleted. Original messages stay in the ledger, and raising the budget later can bring them back. The docs are honest that “summaries are lossy” and older details fade as a conversation runs on.

Retention rules that stop you paying for stale payloads

Retention policies decide what old parts look like when they are replayed. You can say that images are sent once and never again, that results from any tool whose name starts with gh_ are cut to their head and tail after they age, or that failed tool calls are trimmed harder than successful ones. A pruned attachment “stops being sent, and stops being billed, on every subsequent turn.” Truncation markers are on by default because, as the docs put it, “a model that cannot tell output was cut will invent the missing middle.” Whoever wrote that has debugged agents in production.

Namespaces, run records and cost tracking

Namespaces group chats by user or project, and each carries its own budgets, retention rules, system prompts and saved tools, all versioned so you can roll back. Every call to the responses endpoint produces a run record showing what was assembled, which model served it, what it cost and whether compaction ran. If you bill your own users for AI usage, that ledger is half your billing system. Turns are claimed with a durable lease, idempotency keys make a retry safe for 24 hours, and if your process dies mid-conversation the next one picks it up from the chat ID alone.

How Much Does Twigg Cost?

Twigg costs what the model provider charges plus 5%, with no subscription and no monthly fee. The public catalogue states it outright: prices “include the platform margin of 5% over provider cost.” So $1,000 of model spend costs $1,050 through Twigg. The catalogue lists every model with input, output and cache rates per million tokens. One example: claude-sonnet-5 is shown at $2.10 per million input tokens and $10.50 per million output tokens, margin included. Reasoning tokens are not billed twice, and the rates are frozen when a run opens, so a catalogue change never applies backwards.

Billing is prepaid. Your organisation holds a balance in US dollars, top-ups start at $5 through Stripe, and an auto top-up rule can refill it when it falls below a threshold. Before each run, Twigg checks the balance against the input plus the most the output could cost, caps the output to what is left, and returns a 402 error if the balance cannot cover the input alone. New accounts get $5 in welcome credits without a card and another $5 after adding one. That is enough for a real prototype.

For comparison, OpenRouter’s Standard plan takes a 5.5% platform fee and its Business plan 8%, and OpenRouter does not store your conversations. So Twigg charges slightly less than the best-known router and does more for it. The costs it does not show on the rate card are the ones you would expect: compaction is a real model call and appears as its own line on your bill, and there is no way to bring your own provider keys that we could find in the docs. If you have negotiated rates or committed spend with a provider, they do not carry over.

Where Does Your Conversation Data Go?

This matters more for Twigg than for a plain router, because Twigg keeps the transcripts. The security page says production services and storage run on Google Cloud in the United States. Twigg AI Ltd is a UK company, and it says personal data is transferred under the UK’s International Data Transfer Agreement or the EU’s Standard Contractual Clauses. It will sign a data processing agreement on request. It states: “We do not train on your data and we do not sell it.” Request logs capture method, path, status and latency, never bodies. No SOC 2 or ISO certificate is mentioned anywhere we looked.

One claim needs a footnote. The security page says a turn “sends the assembled context to the provider your request named, over HTTPS, and nowhere else.” The context docs say something narrower: “Compaction currently uses OpenAI’s gpt-5-6-luna, regardless of the model selected for the chat turn. History selected for summarisation is sent to OpenAI.” It adds that customers cannot change the compaction model. So a team that picks Anthropic or Google for every turn, perhaps for policy reasons, will still have older history summarised by an OpenAI model once a chat grows long enough. The docs disclose this clearly. The security page should say it too.

You can avoid it. Setting a namespace to drop the oldest history, in place of summarising it, makes no summary calls at all. You lose the memory of early turns, but nothing goes to a provider you did not name.

What Does Twigg Leave Out?

There is no SDK. The docs say so with some pride: “An HTTP client that can read server-sent events is the whole integration.” There is an OpenAPI 3.1 spec for code generators and a GitHub repo with a one-file Node example, but you will be writing your own stream parser and retry logic. Some teams will like that. Teams used to installing a package and calling one function will not.

The scope is narrow by design: “conversational turns with tools”, and “not embeddings, image generation, batch or fine-tuning.” Twigg never executes a tool. It tells you one was called and waits for your result. Only models with a context window of at least 128,000 tokens are supported. And while the pitch is no provider lock-in, your conversation ledger now lives with a young company in place of a large one. The history endpoint returns everything stored, including pruned attachments, so there is a way out. You should still write the export job before you need it.

Who is Twigg Best For?

Use Twigg if you are building a chat product, a support agent or a coding agent where conversations run long, tools return large payloads, and you want to switch models per turn without rewriting history. It suits small teams who would otherwise spend their first month on a messages table and a truncation function. It is also a good fit if you resell AI usage and need per-user cost records from day one.

Skip Twigg if you make single-shot calls with no conversation state, since a plain router is simpler. Skip it if your data must stay in the EU or UK, if procurement requires a security certificate, or if you rely on your own provider contracts and keys. And if an extra network hop and an extra vendor in the path of every model call worries you, that concern is fair. The catalogue does report each model’s success rate over the last 24 hours, but the docs call that “a traffic-based measure, not continuous monitoring or an uptime SLA”, and we found no SLA for Twigg itself.

Best Twigg Alternatives

OpenRouter is the alternative if all you need is one key for many models. Its Standard plan carries a 5.5% platform fee and Business 8%, with a free tier limited to free models. It is stateless, so you still store, trim and replay the conversation yourself. Many teams will pair a router with their own database and never need more.

Letta takes the opposite approach and runs the whole stateful agent for you, memory included. It has a Free plan at $0 limited to three stateful agents with your own model keys, Pro at $20 a month for up to 20 agents, and a developer plan at $20 a month with usage-based pricing and unlimited agents for teams building on its API. Choose it if you want the agent loop managed. Choose Twigg if you want to keep the loop and hand off only the context.

Mem0 is a memory layer that sits beside your own message store. It extracts facts from conversations and retrieves them later. Hobby is free with 1,000 retrieval requests a month, Starter is $19 a month and Pro is $249 a month. It solves long-term recall across sessions. It does not fit a transcript to a window or translate between providers.

Final Verdict: Is Twigg Worth It?

Twigg picks one unglamorous problem and solves it with care. The retention rules are the best we have seen in a hosted product, the billing model is easy to reason about, and 5% is a fair price for deleting a month of plumbing from your roadmap. The docs read like they were written by people who have been burned by context bugs, and they tell you things a marketing team would cut, such as lossy summaries and balances that can dip slightly negative under concurrent runs.

The risks are the usual ones for young infrastructure. It sits in the critical path of every model call, it stores your users’ transcripts in the United States, it has no certificate to show a security reviewer, and its compaction step quietly involves OpenAI whatever model you chose. Start with the $10 in credits and a side project, turn on drop-oldest mode if provider choice matters to you, and keep an export script handy. For prototypes and small production apps we would use it. For regulated workloads, wait. We rate Twigg 7 out of 10.