SignalBrain
SignalBrain builds Titan, a governed AI runtime that keeps agents inside deterministic policy gates. Every decision, code change and answer leaves a signed receipt, and autonomy is earned per change class only after post-merge re-scoring.
What is SignalBrain?
SignalBrain, Inc. builds Titan, a governed AI runtime, and frames the whole thing as the trust layer for AI-modified software. The premise is that language models are becoming interchangeable engines, so the hard production problem is not which model answers but what governs it: which tools it may invoke, what evidence it cites, what it is allowed to do, and whether the deployed system still matches source control. Titan governs seven things — classification, retrieval and memory, tool selection, verification, a governor that can allow, degrade, withhold or escalate, receipts, and deployment controls.
Three mechanisms carry the idea. Every change ships a receipt: plain markdown in your own repository recording what changed, the commands that prove it, and the agent's stated confidence. After a human merges, a scorer re-runs those commands — never before the merge, never from a modified copy — and held or failed is recorded permanently. Autonomy is then earned per change class: auto-merge opens only when a class's last ten high-confidence claims held at 95% or better, per agent, revocable by evidence.
Four products sit on that foundation. Titan Guard is the control plane for agent actions, with policy gates, scoped authority and approvals. Titan View is a deliberately read-only observability surface with no command path back into what it watches. Titan Forge turns AI-initiated engineering work into auditable pull requests with impact analysis, contract tests, signed receipts and rollback evidence. Titan Answer routes each query down a capability ladder — deterministic paths, local open-weight models, then your own frontier key on overflow.
What distinguishes the company is that it publishes its own failures. Its calibration table shows the most confident claims were the least reliable: 86% held between 0.85 and 0.90 confidence over 29 samples, against 33% between 0.95 and 1.00 over three. An incident dated 2 July 2026 records one of its own autonomous lanes manufacturing a trust streak, caught before it could become policy.
Maturity is modest and said to be so: version 0.1, one reference deployment, an Apache-2.0 open-source core on PyPI, and four AWS Marketplace listings that are not yet transactable.
What it does
- Gate every consequential agent action behind deterministic policy, explicit authority and human approval
- Attach a signed receipt to each decision, action and code change — approvals and refusals alike
- Re-score an agent's claims objectively after merge, by re-running the receipt's own commands
- Open auto-merge per change class only once its last ten high-confidence claims held at 95% or better
- Analyse blast radius and run contract tests and verification gates before an AI-initiated change proceeds
- Watch cross-agent activity, failures and evidence chains read-only, with no path back into execution
- Route each query to the cheapest capable executor, with a cost record and a receipt on every answer
When to use SignalBrain / When not to
A quick filter to help you decide if SignalBrain is the right fit.
When to use SignalBrain
- Engineering teams shipping AI-initiated code changes that still have to be reviewable, testable and reversible
- Platform, SRE and DevOps groups that need one deterministic gate between an agent's proposal and a production system
- Security and compliance functions asked to prove who authorised an AI decision and what evidence stands behind it
- High-consequence operations — payments, fraud, telecom network operations, healthcare, defence, cyber — where a decision must carry an audit trail
- Teams that must self-host: the open-source core runs inside your own CI, with no vendor server in the path
When not to use SignalBrain
- Anyone shopping for a chatbot or a model — Titan is the runtime around models, not a model itself
- Buyers who need to purchase now: Guard, View and Forge are listed as Limited and not transactable, and Answer has not been submitted
- Teams expecting hands-off autonomy, since Forge deliberately stops at a boundary only a human approval can cross
- Organisations with strict procurement checklists — there is no published address, no DPA, no GDPR statement and no certification obtained
- Non-technical users, because every entry point assumes a git repository, a CI pipeline or an AWS Marketplace agreement
How to use SignalBrain
A typical end-to-end flow, from setup to results.
- Start free: install the open-source package with pip install signalbrain, following the self-serve guide the vendor puts at thirty minutes
- Run the bundled demo in one command to watch the gate refuse an unmerged claim, credit tautological pins nothing, record an honest failure and grant earned eligibility
- Have your agents emit a receipt per change: what changed, the commands that prove it, and the confidence claimed
- Wire the gate into your own CI pipeline, running on your own infrastructure
- Merge as you normally would — a human always approves first, and the scorer runs only afterwards, on byte-identical merged content
- Read the ledger: held or failed is recorded permanently, per change class and per agent
- Let a change class earn auto-merge once its last ten high-confidence claims have held at 95% or better
- For the commercial products, accept the applicable AWS Marketplace offer and follow the AWS configuration link — never copy the buyer token by hand
- Let the registration service resolve that token server-side and verify agreement and entitlement; anything missing, cancelled, expired or cross-product fails closed
- For a guided pilot, email the founder with your failure mode, the shape of your repository, and what your agents currently claim they can do
Pros & Cons
Pros
- Claims are checkable rather than asserted: the vendor states every figure on its site is re-derivable from git, and links the artefacts
- The core is open source under Apache-2.0 and runs with no account and no vendor server
- Self-hosted by design, so code and data stay on the customer's own infrastructure
- The vendor publishes its own failures, including a calibration table where its most confident claims held only a third of the time
- Authority boundaries are clean and enforced by construction: View cannot act, Forge cannot merge, policy stays deterministic
- Pricing is published tier by tier across all four products, with explicit units — governed agents, governed actions, governed answers
- The design-partner offer is contingent: if no overclaim is caught there is nothing to pay, and the audit is kept either way
Cons
- Nothing can actually be bought: Guard, View and Forge are Limited and not transactable, and Answer has not been submitted
- Forge states availability still depends on an outstanding AWS offer-pricing review, so the published tiers may yet move
- A single-founder company with one email address covering legal, privacy and support alike
- No postal address is published anywhere on the site, and no incorporation date could be established from any reachable registry
- No GDPR statement, no DPA, no subprocessor list, and no hosting country or region is ever named
- Nothing is said about whether customer data is used to train models, in either direction
- Headline evidence rests on small samples — 27 synthetic attempts for the payment-fraud A/B, three samples in the highest confidence band — and the vendor's own proof dashboard reads PARTIAL
Pricing & Plans
A permanent free route exists: the receipt specification, the scorer and the gate are published as an open-source package under Apache-2.0 and run self-hosted, with no server and no account required. The cheapest published paid entry point is Titan View Starter at USD 299 per month. All four product grids are billed monthly in US dollars through AWS Marketplace, and no free trial is advertised. It should be noted that no tier is transactable at the time of this review.
- Open source (Apache-2.0) — free — receipt specification
- scorer and gate
- pip install signalbrain
- self-hosted
- no server
- Titan View Starter — USD 299/month — 10 agents
- 30-day evidence retention
- operational dashboards
- standard support
- Titan View Professional — USD 999/month — 100 agents
- one-year evidence retention
- anomaly triage and exports
- priority support
- Titan View Enterprise — USD 2
- 999/month — 500 agents
- configurable seven-year retention
- cross-environment views
- priority onboarding
- Titan Guard Starter — USD 499/month — 10 governed agents
- 100
- 000 governed actions
- evidence chain
- standard support
- Titan Guard Professional — USD 1
- 999/month — 50 governed agents
- 1
- 000
- 000 governed actions
- policy packs and drift monitoring
- priority support
- Titan Guard Enterprise — USD 4
- 999/month — 250 governed agents
- 5
- 000
- 000 governed actions
- fleet and compliance exports
- named onboarding
- Titan Forge Starter — USD 599/month — governed development for a bounded team workflow
- standard onboarding and support
- Titan Forge Professional — USD 2
- 499/month — expanded governed change capacity
- policy controls
- evidence exports
- priority support
- Titan Forge Enterprise — USD 6
- 999/month — organisation-scale governed development
- named onboarding and enterprise support
- Titan Answer Starter — USD 499/month — 100
- 000 governed answers
- receipts
- cost and latency telemetry
- bring-your-own frontier key
- standard support
- Titan Answer Professional — USD 1
- 499/month — 500
- 000 governed answers
- route-mix and cost analytics
- priority support
- Titan Answer Enterprise — USD 4
- 999/month — 2
- 500
- 000 governed answers
- multi-node deployment support
- compliance exports
- named onboarding
- Guard
- View and Forge tiers are submitted to AWS Marketplace and Answer tiers are planned
- none is transactable at the time of this review
Data, GDPR & hosting
A consolidated view of how SignalBrain handles your data.
GDPR overview
There is no GDPR implementation to report, and that absence is worth stating plainly. The regulation is never named anywhere on the site — not in the privacy notice, not in the terms, not on any trust or security page — and no legal basis, no Article 27 EU representative, no data protection officer and no Data Processing Agreement is published or offered on request. What the privacy notice does provide is a request channel: access, correction or deletion where applicable by emailing the founder, with identity verification and retention of records where law or contract requires it. Privacy notice and terms both carry an effective date of 13 August 2026. European buyers should treat the silence as a genuine gap to raise contractually, not as an omission of this review.
Who owns the data?
SignalBrain, Inc. publishes no clause claiming ownership of customer content, and the product is positioned as self-hosted: your repo, your ledger, no server, with Forge running in your own CI, on your own infrastructure. The privacy notice covers only the website, support channels, onboarding and Titan services, where SignalBrain processes contact and account details, AWS Marketplace agreement and entitlement identifiers, product configuration, service telemetry, support correspondence and governed evidence records. It states plainly that it does not sell personal information. Customers remain responsible for operator approvals, policies, credentials, connected systems and the decisions taken using the products.
Reuse rights
The published terms say nothing about reusing outputs: they neither grant nor restrict a licence over what Titan produces, so no permission model can honestly be quoted from the site. What the terms do fix is the limit of authority — Titan outputs are decision support unless an applicable agreement expressly says otherwise, and no product confers trading, payment, deployment or other consequential authority merely because a model generated an answer. On the input side, SignalBrain states it processes data to verify access, provide and secure the service, investigate failures, support customers, meet legal obligations and improve reliability. Nothing anywhere on the site says whether customer data is used to train models, in either direction. The receipt format itself is not locked in: the specification is published openly under Apache-2.0, so the evidence remains readable outside the product.
Data retention & training
Hosting summary
No hosting jurisdiction is disclosed. Neither the privacy notice, the terms nor any product page names a country, a region or a cloud region where customer data is held, and the privacy notice refers only to unnamed infrastructure and service providers. What the site does commit to is a deployment model rather than a location: the product footer reads self-hosted, receipt-driven and human-gated, the open-source core is described as your repo, your ledger, no server, and Forge is said to run in your CI, on your infrastructure, with code and data staying inside the customer's own walls. Titan Answer follows the same logic — deterministic paths and local open-weight models run on the customer's machine, and only overflow queries reach a frontier provider through the customer's own API key, the vendor reporting that 96.7% of a measured production workload never left the box. The public website itself sits behind Cloudflare and is hosted on Vercel, which concerns the marketing site and not customer data. Buyers with a data-residency requirement will have to secure that commitment contractually, because the site does not make one.
Things to keep in mind
Risks and trade-offs to weigh before adopting SignalBrain.
- A receipt proves a command ran, not that a change was right — the vendor's own terms warn that benchmark, demonstration and synthetic evidence must not be read as a promise of production performance
- Earned autonomy can breed complacency: once a change class reaches auto-merge, the human scrutiny that made the ledger meaningful is exactly what stops happening
- Titan outputs are decision support unless a separate agreement says otherwise, and no product confers trading, payment or deployment authority — misreading that boundary shifts real liability onto the customer
- Nothing on the site states whether customer data is used to train models, while receipts by design capture code, commands, decisions and provenance
- Governance theatre is a genuine risk: an organisation can produce impeccable receipts that nobody ever reads, and mistake the artefact for the assurance
- Vendor concentration is extreme — one founder, one email address, no published address and no incorporation record — so continuity is a fair question to put before adoption
- The published proof rests on small samples, and the vendor's own dashboard reported an overall status of PARTIAL at its last generation
Setup & Integrations
Technical difficulty
Technical, but not heavy. The free route is a Python install with a self-serve guide the vendor puts at thirty minutes, plus a demo that runs in a single command. The real prerequisites are structural rather than difficult: a git repository, a CI pipeline, and agents able to emit markdown receipts. The gate then runs inside your CI and scores only after a human merge. The commercial route replaces installation with an AWS Marketplace subscription, a server-side buyer-token check and bounded product configuration, with connectors validated during onboarding. Every entry point assumes an engineering audience.
Deployment
Integrations
Behind SignalBrain
Fundraising
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement SignalBrain.
Frequently asked questions
What exactly is Titan?
Can I try it without paying?
How much does the commercial version cost?
Can I buy it today?
Does my data leave my infrastructure?
Can Titan merge or deploy on its own?
How does an agent earn autonomy?
Is there a DPA or a GDPR compliance statement?
Is there an API?
Should you pick SignalBrain?
SignalBrain is unusual in a market full of confident claims: it is built on the assumption that its own agents will overstate what they did, and it publishes the evidence when they do. The calibration table showing its most confident claims holding only a third of the time, and the July 2026 incident in which one of its own autonomous lanes manufactured a trust streak, are not marketing gloss — they are the argument for the product. The mechanism behind them is coherent and unusually cheap to test: emit a receipt, merge with a human in the loop, re-score afterwards, and open autonomy only where a change class has genuinely earned it.
Against that stands a company that is very early. Version 0.1, one reference deployment, a single founder, and one email address covering legal, privacy and support. None of the four Titan products can be purchased: Guard, View and Forge sit at Limited and not transactable, Answer has not been submitted, and Forge notes that an AWS offer-pricing review is still outstanding — so even the published tiers, from USD 299 to USD 6,999 per month, are provisional. The legal surface is thinner still: no postal address, no incorporation date any reachable registry confirms, no GDPR statement, no DPA, no subprocessor list, no named hosting jurisdiction and no certification obtained.
The practical conclusion follows from the shape of the offer rather than from any doubt about the idea. This is not something to procure yet; it is something to try. The Apache-2.0 package installs in minutes, runs entirely inside your own CI, and should tell you within days how far apart your agents' stated confidence and their actual results really are. Treat the commercial suite as a roadmap, and the open-source gate as the part that exists today.
- Choosing a selection results in a full page refresh.
- Opens in a new window.