Faros

More than 100 connectors bring in data from coding agents and harnesses such as Claude Code, OpenAI Codex, Cursor and GitHub Copilot, from AWS Bedrock, Azure Foundry and Google Vertex, and from GitHub, GitLab, Bitbucket, Azure DevOps, Jira, SonarQube, PagerDuty and Workday, and custom connectors reach other systems. Faros also acts outside itself: its policy engine pushes approved models, budgets and routing rules to the customer's own AI gateway and harnesses, or enforces them in its own gateway on every request.

Sourcefaros.ai/platformread 2026-11-08

A closed loop runs around the agents. Telemetry from every session feeds the engineering world model, the Time Machine scores alternative routes against completed work, and the winning routes become policies the gateway applies to each new request, choosing the model and harness a task goes to. Faros itself does not plan or run multistep coding work, split a task across agents or coordinate handoffs between them; that work stays with the harnesses the customer already uses.

Sourcefaros.ai/platformread 2026-11-08

The engineering world model is a live graph built from the customer's own systems, joining tickets, agent sessions, token flow, commits, pull requests, CI verdicts and incidents, and it stays current as tools change. Attribution runs per session, pull request, team and model. Agents can connect to Faros to query that graph and build dashboards, and to pull repository specific skills and rules mined from the company's code history, giving them fuller context when they plan and execute.

Sourcefaros.ai/platformread 2026-11-08

Admins decide what agents may use. Policies name which models, harnesses and configurations each team may use, and the gateway stops requests to anything off that list, including tools used in shadow mode. Quotas by team, project or business unit carry a chosen response as they near the limit, and one option is escalating the request to a person for approval. Violation monitoring shows which teams are breaching budget, quota or risk policies while the work runs.

Sourcefaros.ai/solutions/ai-governanceread 2026-11-08

Team level policy is the access model: admins set which models, harnesses and configurations each team may use and how much it may spend, and the same rules bind human builders and agents. API keys can carry an expiration date, and WorkOS and Frontegg, both identity platforms, are among the subprocessors. Faros holds SOC 2 Type II, audited by Insight Assurance with no exceptions, ISO 27001:2022, recertified in 2026 with no nonconformities, a GDPR assessment and a CSA STAR listing.

Sourcesecurity.faros.airead 2026-11-08

For any repository, team or time window, the audit trail lists every AI session, the model behind it, the policy in force and the cost, and it can link a merged pull request to its ticket, the model that wrote the code and what that code cost to produce. Spend attribution, efficiency benchmarking and violation monitoring sit on the same data, and every policy action lands in the trail. Faros also follows whether agent code was merged, reworked or caused an incident.

Sourcefaros.ai/solutions/ai-governanceread 2026-11-08

Faros keeps a durable record of engineering activity, and that record is the product's own data model: sessions, spend, commits and outcomes held for analysis. The skills and rules it supplies to agents are instructions derived from code history. No store holds what an individual agent learned in one session for use in the next, by user, team or workflow, so agent memory stays with the coding tools themselves.

Sourcefaros.ai/platformread 2026-11-08

Three deployment models are on offer: SaaS, a hybrid, or on premises, where the general terms license the software for installation in the customer's own infrastructure or with its cloud provider. The SaaS runs on AWS in separate US and EU1 regions, each with its own API and connector addresses. Faros says customer data never leaves the customer's boundary, and policy enforcement can run through Faros's gateway or a gateway the customer already operates.

Sourcefaros.ai/general-termsread 2026-11-08

No prebuilt agents come with the platform. What Faros hands over is derived from each customer's own history: model routes proven by the Time Machine, repository specific skills and rules for agents, and SDLC findings, each delivered ready to apply. Around that sit ready made measurement frameworks, among them a field guide of three outcome signals and eleven guardrail metrics, developer productivity dashboards and a public Route Index comparing model routes.

Sourcefaros.ai/platformread 2026-11-08

Two kinds of event set Faros in motion. Every coding agent request that passes through the gateway is checked against policy and routed as it happens, and a team approaching its quota sets off the response an admin chose in advance: throttling lower priority work, rerouting it to a cheaper model, or escalating it for approval. Data from the connectors arrives continuously. No schedule, webhook or inbound queue starts a piece of work on its own.

Sourcefaros.ai/solutions/ai-governanceread 2026-11-08

Admins choose the models. Routing policies send each kind of task to a specific model and harness, such as Claude Opus, Sonnet or Haiku by task complexity, approved model lists are set per team, and new models run against the same historical tasks before a route changes. Faros works with any model, harness or router, covers AWS Bedrock, Azure Foundry and Google Vertex, and enforces routes through its own gateway or the customer's existing one.

Sourcefaros.ai/solutions/optimal-model-routingread 2026-11-08

The Faros API runs in each region from its own addresses and takes API keys that can be set to expire, and an API reference sits beside the guides and recipes in the documentation hub. Agents can connect to Faros directly to pull data, build dashboards and fetch repository specific context. Custom connectors extend ingestion to any other system, and Faros Community Edition, an open source edition with its own documentation site, sits beside the commercial platform.

Sourcedocs.faros.airead 2026-11-08

The Time Machine is an evaluation engine built on the customer's completed work. It replays finished tasks from the organization's own repositories across alternative model and harness routes and scores each on quality, cost and runtime, so a route is proven before it changes in production. Build Your Own Bench turns shipped work into a private benchmark, and new models run through the same tasks as they arrive. On 211 of its own tasks Faros reports roughly half the cost per task at equal or better quality.

Sourcefaros.ai/solutions/optimal-model-routingread 2026-11-08

Connectors, APIs and the AI gateway in the request path are how the platform works. It measures and steers the model calls coding agents make, with no hosted browser, no desktop control and no agent of its own that clicks through web interfaces. When the customer's coding agents use a browser, that activity reaches Faros only as sessions and token spend in the record, and Faros never drives the interface itself.

Sourcefaros.ai/platformread 2026-11-08