MCP GUARD

Agent-fleet control plane

Ship more with agents.Spend less. Prove everything.

Coding agents multiply your output — and your spend, and your audit surface. MCP Guard is the control plane that keeps all three honest: fleets governed by policy, costs priced to the cent, and every line provable — across every runtime you run.

Open dashboardContact sales
built-in gates + plugin SDK
11built-in gates + plugin SDK
ledger, re-verified hourly
SHA-256ledger, re-verified hourly
per-session attestation
Ed25519per-session attestation
fleet spend precision
$0.01fleet spend precision

Governs Claude Code · Cursor · Devin · Copilot · in-house agents

For engineering teams

Run it yourself →

You already run agents. Get the most out of them without the security risk — the control plane, on your repos, today.

For everyone else

Have it built for you →

Need software, not a platform? Our governed fleet builds it — faster and more accountable than a dev shop.

The thesis

The IDE was built for writing code. This is where you ship it. The center of gravity is moving — out of the editor, into the control plane. The terminal and the IDE become where agents work. This is where you do.

01

Typing → Delegating

One engineer used to mean one cursor. Now it means a fleet of agent sessions running in parallel — and the constraint stops being how fast anyone types.

02

Reviewing → Policy

You cannot eyeball ten thousand diffs. You can write the rules once — approved agents, models, MCP servers, tools — and have every merge evaluated against them, mechanically.

03

Trust → Proof

“An agent wrote this” satisfies no auditor. A hash-chained ledger of every decision, exportable and independently verifiable, does.

The control plane

Govern. Observe. Intervene. Prove.

Not a CCTV wall for watching a thousand agents — an exception-routed surface. Set policy once, let the fleet run, touch only what the policy escalates, and keep cryptographic evidence of all of it.

Govern

/policy

Policy as code, simulated before it ships.

  • Eleven built-in gates — test coverage, schema discipline, PII scanning, agent attestation, MCP-server allowlisting, model policy — plus a sandboxed plugin SDK.
  • Replay the last 30 days of real agent activity against a draft policy before publishing. Nobody else has the data substrate to do this.
  • Graduated cost limits per session, per actor, per agent type — breach notifications fire exactly once.

Observe

/fleet

Every in-flight session, every runtime, one surface.

  • Live via SSE, attested via Ed25519 — Claude Code, Cursor, Devin, Copilot, and in-house agents in the same view.
  • API spend priced per segment and verified to the cent, with ranked optimization insights and one-click model write-back.
  • Stuck sessions escalate through warn → alert → flag tiers; abandoned work can't hide.

Intervene

/inbox

Exception-routed triage, not a CCTV wall.

  • Gate failures, attestation misses, cost breaches, and policy denials land in one queue — most sessions never need a human at all.
  • Resolve, snooze, assign, waive — every action writes a hash-chained audit row with the operator's reasoning.
  • Moot exceptions auto-resolve when the PR ships; the queue stays about decisions, not housekeeping.

Prove

/audit

A ledger your auditor can verify alone.

  • Every policy decision and agent action chains by SHA-256 in Postgres; the chain re-verifies hourly.
  • CSV / JSON / PDF evidence packets on demand, scoped by repo, gate, actor, and time range.
  • PIN-gated, time-bounded auditor shares — never raw database access, never an opaque PDF.

What it looks like

Three surfaces. One control plane.

Mock screenshot of the MCP Guard fleet view: a table of in-flight agent sessions across Claude Code, Cursor, Devin, and an in-house runtime, with state chips and a live SSE indicator.
/fleetEvery in-flight agent session — across every runtime — in one operator surface. Live via SSE, attested via Ed25519.
Mock screenshot of the MCP Guard exceptions inbox: a list of policy violations grouped by type with severity chips and suggested resolutions.
/inboxPolicy violations land in a triage inbox. Resolve, snooze, dismiss — every action writes a hash-chained audit row.
Mock screenshot of the MCP Guard policy simulation view: a draft policy diff on the left, a replay outcome on the right with per-rule block counts.
/policy/simulateReplay the last 30 days of agent activity against a draft policy before you publish. See what would have blocked, and why.

How it works

Three steps from install to auditor-ready.

  1. 01

    Connect

    Install the GitHub App, point your agent runtimes at the /v1/agent-runs endpoint, and pick a policy template. First check-run in minutes.

  2. 02

    Govern

    Every PR runs the gates. Every session is attested and costed. The merge button becomes the policy boundary, not a hope.

  3. 03

    Attest

    Auditors get a signed-URL share into the hash-chained ledger. Compliance reviews evidence; engineering keeps shipping.

For delivery firms

A new operating model for consulting: sell outcomes, run fleets, prove everything.

The leverage of a consulting firm used to be headcount. Now it's a governed agent fleet with a senior engineer at the controls. We run our own delivery practice on MCP Guard — opportunity sourcing through client handoff — and license the same platform and playbook to firms making the shift.

  1. 01

    Source

    A read-only scout agent walks your saved searches and lands qualified opportunities in the Hunter pipeline.

  2. 02

    Engage

    Quote, discovery workspace, and kickoff brief in one engagement record — with a budget guardrail watching cost-to-deliver from day one.

  3. 03

    Deliver

    A governed agent fleet does the work inside a provisioned client repo, seeded with a PR-reviewed, versioned, licensed agent pack.

  4. 04

    Hand off

    Deliverables tracked against acceptance, a client status portal behind secure tokened shares, and audit-grade evidence of everything that ran.

What your client sees

A live status portal — deliverables against acceptance criteria, progress without meetings, behind a secure time-bounded link. Self-serve intake with secure document upload. Evidence, not vibes.

What your firm sees

Realized margin per engagement — fixed fee against cost-to-deliver, priced from the same to-the-cent fleet telemetry. Budget guardrails at 70 / 90 / 100%. A portfolio control tower across every engagement.

Need software built? See the consulting practice →

Instrumentation

Everything above is shipped. So is all of this.

01

Cost limits

Graduated thresholds per session, UTC-day, actor, or agent type. Worker-evaluated — breaches notify exactly once, even if nobody opens the dashboard.

02

Policy simulation

30-day replay of real agent activity against any draft. See exactly what would have blocked before you publish.

03

Optimization recs

Ranked cost and speed recommendations with one-click model write-back — and realized-savings tracking after you apply them.

04

Agent topology

Declared agent team versus the invocation graph that actually ran, cross-session, with gap and drift overlays.

05

Stuck-session escalation

Three configurable tiers — warn, alert, flag — with per-stretch re-arming and snooze. Terminal flags mark work abandoned.

06

Evidence export

CSV, JSON, and PDF packets from the hash-chained ledger; PIN-gated auditor shares with email allowlists and trust banners.

07

Plugin gates

Write custom gates against a typed SDK; they run in a hardened subprocess sandbox, never inside the engine.

08

Seven-role access control

Delivery, compliance, sales, and operator seats with server-enforced route floors — pending invitees see nothing until seated.

What auditors see

A hash-chained ledger, exported on demand.

Every policy decision and agent action is written into a Postgres ledger whose rows chain by SHA-256. The chain is verified hourly. Reviewers get a PIN-gated, read-only, time-bounded share of the slice you authorise — never raw database access, never an opaque PDF you have to vouch for in person.

Run your fleet like it's production.Because it is.

Open dashboardContact sales