Privacy & data

Your code runs through our servers. Here's exactly what we do with it.

Talon sits in the path between your coding agent and the model. To add the right context and measure what it saves you, your prompts, the model's responses, and a structural graph of your code are processed on Cogmeta's infrastructure. This is the plain-English version of what that means.

Last updated · 31 August 2026Version · alpha-launchReading time · ~7 min

The short version. In twelve seconds.

Talon is a hosted service. Your agent's traffic is routed through our proxy and a graph of your code is built and stored on our servers, because that is how the product adds context and measures cost. In exchange we make hard commitments: we never train on your data, we never sell or share it, retention is bounded, and access is least-privilege and logged.

The table on the right is the honest summary — what we process, and the limits we hold ourselves to.

What we process
Prompts, model responses, and a structural graph of your code.to inject context & score cost
Where it runs
Through Cogmeta's proxy; graph stored per-tenant on our infra.not on your machine only
Training
Your code and prompts are never used to train any model — ours or a provider's.contractual · no exception
Selling / sharing
Never sold, never shared with advertisers or data brokers.no exception
Hosted
AWS, multi-region.multi-AZ · per-tenant isolation
Self-hosted
Single-tenant / self-host to keep data in your boundary — on the roadmap.
What actually moves

The data flow, in three frames.

No policy language, no ambiguity. A literal description of what passes through Cogmeta during a typical session.

Passes through our proxy

Session traffic

  • Your prompts to the model
  • The model's responses
  • Source and diffs included in that context
  • The context Talon injects
→
Stored on our servers

Your code graph

  • File paths & symbol names
  • Structural relationships between them
  • Anonymous usage counters
  • Metadata derived from your source at index time — never the source itself
→
What we never do

The limits we hold

  • Train any model on your code or prompts
  • Sell or share your data
  • Build cross-customer profiles
  • Store your git remote credentials
  • Use tracking cookies or fingerprints

Want your code and prompts to stay inside your own boundary? Single-tenant and self-host options are on the roadmap.

Data we process

Four categories. If we add a fifth, it will appear here first and a notice will go out before it takes effect.

  • Session traffic (prompts & responses). Because Talon proxies your agent's calls to the model, your prompts and the model's responses — including any source, diffs, or context they contain — pass through our servers so Talon can inject structural context and measure cost. They are processed to deliver the service and are not used to train any model or sold to anyone. Retention of request and response bodies is covered under Retention & deletion.
  • Code graph. A structural index built from your source at index time: file paths, symbol names, and the structural relationships between them — never the source itself. Stored per-tenant on our infrastructure.
  • Session counters. Turn count, token counts by category, latency percentiles, error codes. Used for the attribution dashboard and our own SLO reporting.
  • Account data. Name, email, workspace, role, billing contact — what you gave us at sign-up and what we've added as you use the product.
You can see what Talon sends on any request: run talon debug proxy --tail in a project to watch the exact traffic that transits our proxy, and talon debug graph to inspect the structural graph built from your code.

How we use it

For a small, fixed set of purposes — each tied to a specific lawful basis under UK GDPR Article 6:

  • Score context · contract (necessary to deliver the product you signed up for)
  • Route model calls · contract
  • Render the attribution dashboard · contract
  • Produce your invoice · contract + legal obligation (tax records)
  • Troubleshoot and operate the service · legitimate interests — we access content only transiently, only as far as needed to detect and fix a fault, and never reuse it for anything else
  • Improve the service in aggregate · legitimate interests — using only system-generated metadata (usage counters, latency and error telemetry), never the content of your prompts, responses, or code

The distinction that matters: your content — prompts, responses, and the code inside them — is processed only to deliver the service to you. We never use your content to train models (ours or anyone else's) or to improve our product; anything we learn from to make Talon better is our own system-generated metadata, not your content. We do not sell or share data with advertisers, data brokers, or analytics networks, and we do not build profiles across customers.

Cookies & tracking

We don't use cookies for tracking, advertising, or analytics. Session cookies for sign-in only — strictly necessary, no consent banner required under UK PECR.

Model providers

Talon proxies your agent's requests to third-party frontier model providers on your behalf. Because the proxy is in the path, your prompts and the provider's responses transit Cogmeta before and after they reach the provider. We route only to the provider a request is destined for, and we enforce zero-retention mode on every outbound call where the provider offers it.

You reach a model one of two ways. On Cogmeta credits, we hold the account with the provider and bill you for usage. Bring-your-own-key (BYOK), where offered, forwards the request under your own provider account and key — this changes who holds the upstream relationship, but the request still transits our proxy, so it does not remove Cogmeta from the data path.

Provider zero-retention means your prompts and responses are not retained by the provider beyond the API call, and are never used to train their models. How Cogmeta itself handles that traffic is covered under Data we process and Retention.

Sub-processors

Two categories. The live, authoritative list is maintained here; adding one is a sub-processor change (30-day notice per DPA terms), not a change to this policy.

  • Infrastructure. Amazon Web Services — hosting of the proxy and your code graph (eu-west-2 / London, multi-AZ).
  • Model providers (frontier LLM vendors). The vendors we route inference to on the credits model. Today: Anthropic. As we add model support — for example OpenAI, Google, or xAI (Grok) — the new vendor is listed here under the same terms: routed through our proxy, zero-retention enforced where offered, no training on your traffic. Under BYOK the provider you configure is your own sub-processor, not ours.

As we add other subprocessors (payment processing, email, etc.), they'll appear here first. Changes get a 30-day notice per DPA terms.

Residency & transfers

Today, all metadata is hosted in AWS eu-west-2 (London). Customers requiring a specific region (US, APAC) can talk to us — multi-region is on the roadmap. EU/UK transfers use Standard Contractual Clauses where required by the contract.

Retention & deletion

Defaults — all configurable on Team and Enterprise tiers:

  • Prompt & response bodies (proxy traffic) — processed in transit to serve the request; retained only as long as needed to operate and debug the service, then deleted. Not used for training, never sold.
  • Code graph — until you delete the project · incremental
  • Session counters / operational logs — 30 days · adjustable
  • Audit logs — 12 months default
  • Billing records — 13 months · UK statutory minimum

Account deletion triggers a 30-day grace period, after which everything except legally-required records is irrecoverably erased. Erasure proof is available on request.

This retention table is the single source of truth for how long we keep each data category.

Security practices

TLS 1.2+ everywhere in transit. Encryption at rest using AWS-managed KMS keys. IAM least-privilege with logged operator access. Per-tenant logical isolation; cross-tenant access is structurally impossible.

Audited compliance (SOC 2 Type II, ISO 27001) is on the roadmap, not in place today. We'll publish realistic timelines as they firm up rather than claim certifications we don't hold.

Age & children

Talon is a developer-tooling product intended for adults building software. We do not knowingly collect data from anyone under 16. If you believe a minor has signed up, email [email protected] and we'll remove the account within 72 hours.

Contact

Questions, requests to exercise your rights, or just to say hello:

  • General & data requests — [email protected]
  • Security disclosure — [email protected]
  • Cogmeta Ltd. — UK-registered, Companies House 17078018. Registered office: 71-75 Shelton Street, London WC2H 9JQ. Service-of-process address available on request via the contact form.

Data Protection Officer. No formal DPO required at our scale; data-subject queries route to [email protected]. Article 30 Record of Processing Activities available to data-subject regulators on request.

Your rights

Six requests we'll always honour.

01 / GDPR Art. 15

Right to access

Download everything we hold about you as a signed JSON export. Delivered within 7 days.

02 / GDPR Art. 16

Right to rectify

Correct any inaccurate personal data directly from the account page or via email request.

03 / GDPR Art. 17

Right to erasure

Delete your account and all personal data. 30-day grace period, then permanent.

04 / GDPR Art. 18

Right to restriction

Pause processing while a dispute is resolved. Your workspace continues; analytics pause.

05 / GDPR Art. 20

Right to portability

Export your scoring graph and configuration in an open format, importable anywhere.

06 / GDPR Art. 21

Right to object

Opt out of aggregate telemetry. Product continues with graceful degradation of tuning.

Californian residents have the same rights under the CCPA, plus a right to know sales & shares of personal information. We have none to report — we do neither.