Know what your agents do.
Keep control of what happens next.
Видьте всё, что делают ваши ИИ-агенты, блокируйте недопустимое и сокращайте расходы. Nexus фиксирует каждое действие, применяет политики в реальном времени и отслеживает каждого агента, идентичность и соединение.
Install: npm i -g @swfte/nexus. Runs on macOS and Linux; Windows is coming.
03 SEE IT WORK
See Nexus at work.
Overview
One record of what your agents do.
Sessions, rule matches and recorded spend land on one page, and every figure names the source it came from. When a source goes quiet, the gap is drawn as a gap, never as a zero. Pick a focus to see what needs a decision first.
01 Developer activity
02 Capture & policy
03 Shared evidence
From the local session to the enterprise decision.
Close to the work.
Clear across the estate.
Console, DevTools and the desktop workspace are parts of one family. The role differs; the evidence connects.
See the estate. Find the exposure. Direct the response.
An enterprise view of agents, identities, connected systems and developer activity. Investigate findings, inspect policy decisions and understand the evidence behind reported costs.
- Agent and identity inventory with relationship and exposure views.
- Findings, policies and pending approval records in a shared governance surface.
- Token attribution and application cost views that distinguish reported usage, derived prices and missing evidence.
- Shared Swfte sign-in and an authenticated backend proxy are implemented in current source.
Integration availability and boundaries
- Data availability depends on connected producers; missing cost and governance history remain absent.
- The current canonical console is /v2. Older /nexus, /v2c and /v2-legacy routes redirect; some retired surfaces have no modern equivalent.
- Cloud availability, tenant configuration and service-level terms are agreed per deployment. Swfte holds SOC 2 Type II and ISO 27001 certifications; see the security page for scope.
Make every coding-agent session accountable.
Capture activity locally, apply policy at supported execution hooks and inspect the cost of AI-assisted development. Start with Claude Code or Codex and choose what your workspace receives.
- Wrap Claude Code and Codex sessions through the Nexus CLI.
- Local event ledger with a metadata-only default and configurable privacy tiers.
- Policy checks for supported tool actions, protected paths and configured restrictions.
- Session, change and usage evidence with savings instrumentation where a baseline is available.
- Optional workspace forwarding and a local-session mode that disables captured-data forwarding.
Integration availability and boundaries
- Provider registry entries are not equivalent to installed hooks or complete enforcement coverage.
- Codex coverage can be partial; shell integration can fail open unless strict mode is configured.
- No universal savings percentage, guaranteed cost reduction or automatic quality guarantee is established.
- Local-only captured data does not mean the wrapped model itself makes no network requests.
- The open-source/self-hosted stack and pip distribution are described as forthcoming in the README; no general-availability claim is warranted.
One workspace for the agents doing the work.
Run and inspect coding sessions in a native workspace, with visible governance coverage, session context and a runtime designed to preserve continuity across tools.
- Native Tauri application with a Rust session daemon and interactive attached terminals.
- Session fleet view with start/end controls backed by native commands.
- Governance states distinguish governed, ungoverned, partial and unknown coverage.
- Cost components distinguish priced, partial and unknown spend.
- The Baton contract defines work handoff and lineage between harnesses; role and workflow design share that boundary.
Integration availability and boundaries
- Desktop release availability, supported installation targets and commercial packaging were not established in this source audit.
- Broad multi-harness positioning does not prove equal hook coverage across all named agents.
- Roles, workflows and handoff architecture must not be presented as fully shipped desktop journeys solely because design documents exist.
- The root README foundation label is incomplete: the current tree contains substantial Rust and native application implementation.
[ 02.A – 02.L ] INSIDE NEXUS
Inside Nexus, screen by screen.
Twelve chapters, each a real screen from the console, assembling as you scroll.
N1 [ 02.A ]Overview
One record of what your agents do.
Sessions, rule matches and recorded spend land on one page, and every figure names the source it came from. When a source goes quiet, the gap is drawn as a gap, never as a zero. Pick a focus to see what needs a decision first.
Available- 01Key numbers
- 02Is the data arriving?
- 03Gaps stay gaps
- 04Pick a focus
N2 [ 02.B ]Sessions
Follow one session from prompt to change.
Every coding-agent session is listed with who ran it, where, and whether it is waiting on a person. Open one to read its ledger: each event as recorded, with what it touched, and the record's limits stated where it is thin.
Available- 01Live now
- 02Open a session
- 03What it did
- 04Every event, as recorded
N3 [ 02.C ]Rule matches
Know what each rule actually did.
Every match is recorded as blocked, ran anyway, observed or not reported, with the rule, the person and the target. A table shows what each rule did, and each device reports whether it enforces, audits or is switched off.
Available- 01What the rules did
- 02Recent matches
- 03Each rule
- 04Across devices
N4 [ 02.D ]Live control
A person decides, and the console says when.
Approve or refuse a risky action with a reason that is stored. Instruct, pause or grant a time-boxed power from the same page. Each request is queued, and the session collects it on its next step.
Available- 01Action waiting
- 02Reason, then allow
- 03Instruct or pause
- 04Time-boxed power
N5 [ 02.E ]Cost
Spend you can trace to a number.
Follow cost from model calls to a charged figure, rung by rung. Measured numbers stay apart from calculated ones, and every saving is labelled by how it was worked out. Where a figure cannot be known, it says so.
- 01Request to number
- 02Measured or calculated
- 03Savings by method
- 04Who spent it
N6 [ 02.F ]Reach
Start from one key. See what it reaches.
Pick the riskiest account or key and follow the links it has to other agents, apps and grants. The console also says what the tree cannot show: how far it looked, what it only named, and that no total is reported.
Available- 01Find the key
- 02Tree
- 03Routes
- 04What it can't show
N7 [ 02.G ]AI discovery
Find the AI. Say how complete it is.
Connect a supported AI provider, run a scoped scan and review what each agent is configured to use. Every scan reports its own coverage, and a person records who owns each agent and why it exists.
- 01Connect a provider
- 02Run a scan
- 03Review access
- 04Business context
N8 [ 02.H ]Running services
Which services run an agent, and which don't.
Every service that reported itself appears with what it measured. Services someone wrote down, and nothing has ever reported, are drawn differently. A service that runs while no application claims it is the one thing marked in the accent.
Available- 01What is running
- 02Gone quiet
- 03Written down
- 04Claimed by nobody
N9 [ 02.I ]Review
One queue. And what each decision rests on.
Rule matches, findings and approvals share one list, ordered by severity then age. The console marks which decisions are stored on the server and which exist only in this browser, and says what is still missing.
- 01One queue
- 02Mark seen
- 03By source
- 04Who decided
N10 [ 02.J ]Session safety
Find secrets in AI sessions, on your machine.
A local scan reads saved Claude and Codex sessions for keys, passwords and personal data. Only a masked report leaves the device, and each finding says how to find it again and what to do next.
Available- 01Scan on device
- 02Masked report
- 03Review a finding
- 04Locate it locally
N11 [ 02.K ]Across industries
Same console, different questions.
Pick a sector. Each one asks something different of its coding agents, and the same screens answer it: what ran, what a rule did, who decided, what it cost, and what could not be measured.
N12 [ 02.L ]Why Nexus
Evidence you can show a reviewer.
Nexus shows what happened, what it cost, what it can reach and who decided, and says plainly where it cannot know. Start with visibility, then agree the controls.
- 01Eight pieces, one console
- 02One chain, laptop to service
- 03Explore each piece
Illustrative demo
Where a figure is estimated, discovery is partial or a decision is kept only in the browser, the console says so on the screen.
Every agent action,
captured at the source.
Nexus wraps the agent runtime itself. Events stream from the terminal to a durable local ledger, then roll up to the governance console: who did what, in which repo, on which terminal.
| Event | When | What Nexus records |
|---|---|---|
| session | Agent session starts | Terminal ID, user (git email), repo, branch, commit, model and provider |
| prompt | Every prompt submitted | Fingerprint and length only at the default privacy tier. No prompt text leaves the machine |
| tool_action | Before every tool call | Tool, target, and a blocked flag with the policy reason when Nexus says no |
| dependency_install | npm / pnpm / yarn / pip / mvn / cargo | Package and version, giving supply-chain visibility for every agent-initiated install |
| file_change | Every edit or write | Repo-relative path plus a protected flag when a policy-guarded file is touched |
| token_usage | Every turn | Input, output, baseline vs. compressed tokens, savings, and cost |
Privacy by default: at the standard capture tier, prompt content is never stored — only a fingerprint and length.
Governance
that pays for itself.
The same event stream that powers the audit trail meters every token. Nexus attributes AI spend per terminal, per repo, and per user, and measures exactly what context compression and output filtering save you.
Baseline vs. actual
Every turn records baseline tokens, compressed tokens, and the delta, so savings are measured, not estimated. You see the real cost of every session and what it would have cost without optimization.
Cost attribution
Spend rolls up by user, repo, terminal, and model. Runaway sessions and expensive workflows surface immediately instead of at invoice time.
Waste elimination
Per-command output filtering and context compression cut the tokens agents burn on verbose tool output: typically the largest single source of avoidable agent spend.
Trace every agent.
Every connection. Every change.
Observability tells you what happened. Nexus also decides what is allowed to happen, and maps the identities and connections behind it all.
In-flight policy blocking
Nexus sits between the agent and your system. Protected files, forbidden commands, and unapproved dependency installs are denied before they execute, with the reason logged to the audit ledger.
Audit trail: prompt → change → why
Every change an agent makes is traceable back through the tool action and the prompt that caused it. Security review stops being archaeology; the trail is already assembled.
Shadow-AI detection
Unmanaged agents, unsanctioned model providers, and unknown non-human identities surface as findings. This is the AI usage your existing SaaS security tooling cannot see.
Identity graph + blast radius
Every agent and non-human identity in one inventory, with a connection graph that answers the question that matters in an incident: if this credential is compromised, what can it reach?
Supply-chain visibility
Agent-initiated package installs are first-class events. Know which agent added which dependency, in which repo, on which terminal, before it ships.
Policy-as-code
Governance rules are declared, versioned, and evaluated continuously, rather than captured as screenshots of settings pages. Findings show exactly which policy failed and where.
Start where the work is.
Broaden when the evidence is.
Start with visibility.
Connect supported developer workflows and establish what is captured, what is missing and what is attributable.
Agree the controls.
Choose the policy boundaries and approval responsibilities that are appropriate for the environment.
Expand the evidence.
Review coverage and adoption before broadening the rollout. Keep the gaps visible.
Go deeper.
No email wall.
Ungated guides on the problems Nexus was built for: no email wall, no PDF download.
Claude Code security
Securing and governing Claude Code in the enterprise.
↗MCP security best practices
Hardening Model Context Protocol servers and connections.
↗Agent observability
Why observing agents differs from observing LLM calls.
↗LLM observability
Tokens, cost, latency, quality, and traces for LLM apps.
↗Shadow AI
What it is, why it is exploding, and how to detect it.
↗AI agent governance
Inventory, policy, audit, and enforcement for agents.
↗Non-human identity
Agents, service accounts, and keys as first-class identities.
↗Zero trust for AI agents
Never trust the agent by default; verify every action.
↗Error encyclopedia
Infrastructure and AI-stack errors explained and fixed.
↗Before you
make a start.
What is Swfte Nexus?
Nexus is an AI-agent governance and observability platform. It captures what coding agents like Claude Code actually do: every session, tool action, file change, dependency install, and token spent. It enforces policy in-flight by blocking violations before they execute, and gives security teams an inventory of agents and non-human identities with an identity graph and blast-radius analysis.
How does Nexus capture agent activity?
A lightweight wrapper hooks the agent runtime (for Claude Code, via native hooks) and streams structured events to a local collector and durable ledger. Events flow to the Nexus console for per-terminal, per-repo, per-user visibility. At the default privacy tier, prompts are captured as fingerprint and length only. No prompt text is stored.
Can Nexus block an agent action, not just log it?
Yes. Nexus evaluates policy before the tool call executes. Protected files, denied commands, and unapproved installs are blocked in-flight, with the reason recorded in the audit trail. That is enforcement, not just observability.
How does Nexus reduce AI spend?
Nexus meters token usage per turn (input, output, baseline versus compressed) and attributes cost per terminal, repo, and user. Teams use it to find runaway sessions, enforce budgets, and measure what context compression and output filtering actually save.
Does Nexus work with agents other than Claude Code?
The capture wrapper is agent-agnostic by design; Claude Code is the deepest integration today via its hook system. The governance console inventories any agent or non-human identity reachable through its connectors, and transcript import covers agents without native hooks.
Is Nexus open source?
Nexus is open-core. Capture, forwarding, and enforcement are open source and always available. The intelligence layer, covering audit trails, metrics, knowledge harvesting, and security modeling, runs against an Enterprise backend: either Nexus cloud or self-hosted.
Implementation evidence is based on current Nexus repositories. It does not establish commercial availability, universal agent coverage or any deployment-specific security guarantee.