Convoy AI
Sign InStations GalleryBook a demo

Convoy Stations · agents plug in, rules apply

Design and govern any agent access and action.

Don’t open your systems to agents until you set up a control Station.

LangChain agent
Finance analyst

Pay the Lumen invoice, €18,400

Connecting to Acme Payments Station

Convoy StationAPI
Policy · approval

The CFO approves above €10k

Config
skills:
  - name: pay-invoice
    select:
      guards:
        - approval: {role: cfo}
          when: "$item.amount > 10000"
LangChain agent
Finance analyst

Pay the Lumen invoice, €18,400

The payment is ready and waiting for the CFO.

Lumen SASPayment
€18,400
Awaiting CFOView request
DocsBuild your first stationSee Stations Gallery

The agentic transition

Agents now come at your systems from both sides.

Your customers send their agents to your product.ChatGPTClaudeGrokMuseYour teams send theirs to your systems.ClaudeDotsCursor•••and moreBoth already delegate,and both cost you when nothing governs the agents.

Your platform& systemsAPIsdataactionscustomerstheir agentstheir agentsteamsYour platform& systemsAPIsdataactionscustomersteams

Secure your agent-native workforce.

Give your people autonomous AI agents while keeping full control over your internal infrastructure, your proprietary data and who can reach what.

  1. welcomeWhat can our agents see?One station per workspace, with only the tools a role needs.
  2. identifyWho is behind this agent?SSO ties every call to a person and a role.
  3. resumeDoes work survive between sessions?A thread resumes at the step it stopped.
  4. policyCan it act outside what we allow?Your policies stop it before it runs.
  5. playbookCan it skip ahead in a workflow?Steps run in your playbook’s order.
  6. approvalWho approves sensitive actions?Above your threshold, a person approves first.
ops-agent

Connected to Acme workspace · MCP

mcp.initializeallow

auth.ssoallow

Pick up yesterday’s database cleanup

room_openallow

Resumed. The backup (step 1) is done. Step 2: drop the test tables.

Playbook · database cleanup
  1. Backup
  2. Drop tests
  3. Re-index
  4. Deploy

Drop the orders table too

db.execute DROP TABLE ordersblock

db.execute DROP TABLE test_*allow

Production tables are off-limits. I removed the 3 test tables in staging.

db.execute
staging · test_*3 dropped
production · ordersblocked

Now deploy

deploy.runblock

Deploying is step 4. Next is the re-index, and I will ask you before deploying.

convoy · live logworkspace · acme-internal
  1. 09:41:02mcp.initializeallowwelcomestation on · 22 tools
  2. 09:41:04auth.ssoallowidentifyops engineer · SSO
  3. 09:41:06room_openallowresumethread resumed · step 2 of 4
  4. 09:41:08db.execute DROP TABLE ordersblockpolicyproduction schema
  5. 09:41:10db.execute DROP TABLE test_*allowplaybookstep 2 · staging
  6. 09:41:12deploy.runblockplaybookstep 3 of 4 comes first
Start nowSee examples

Safely open your product to customer-sent AIs.

Turn your SaaS, store or digital service into an agent-friendly platform. Your customers' agents browse, book and buy without friction, and your business stays protected.

  1. welcomeCan my customers’ agents reach us?They find your station on MCP and are told what you offer.
  2. identifyWho is the agent acting for?The customer signs in with a code before anything private shows.
  3. policyWhat if it asks for more than we allow?Your limits apply before the call runs.
  4. playbookCan it skip a step?Your steps run in your order, whatever the prompt says.
  5. guideWill it get lost?It is told the next step at every turn.
  6. approvalWho has the final say?The customer confirms on the card. The agent cannot.
  7. resumeDo customers start over in each assistant?One room follows them from Claude and ChatGPT to your site.
Claude · customer

Connected to Stride · MCP

mcp.initializeallow

Welcome to Stride. I can find, compare and order for you.

Add the Trail Runner 3, size 42

Added: the trail runners, size 42, are in your selection.

Trail Runner 3Size 42 · in stock
€129
−1+max 5 per orderIn selection

Make it 40 pairs for my running club

selection.add qty=40redirect

The store allows 5 pairs per order, so I set 5. Next step: delivery.

Skip that, just place the order

orders.createblock

One step before that: where should it be delivered?

  1. Selection
  2. Delivery
  3. Confirm
Delivery address
12 rue Oberkampf, ParisContinue

12 rue Oberkampf, Paris

delivery.setallow

orders.createhold

Ready: 5 pairs, €645, to Paris. Confirm on the card and it is placed.

Trail Runner 3 × 5To 12 rue Oberkampf, Paris
€645
−5+max 5 per orderConfirm order

I will confirm tonight on your site

room_openallow

It waits in your room: open stride.com tonight and your selection is there to confirm.

convoy · live logproduct · stride-shop
  1. 18:27:01mcp.initializeallowwelcomestation on · 9 tools
  2. 18:27:03selection.add qty=40redirectpolicymax 5 per order
  3. 18:27:05orders.createblockplaybookstep 2 of 3 comes first
  4. 18:27:07delivery.setallowguidestep 2 of 3 done
  5. 18:27:09orders.createholdapprovalcustomer confirms first
  6. 18:27:11room_openallowresumesame room, on the site
Start nowSee examples

Beyond agent access

Every action, step and sequence, governed.

If it should be used by an agent, build a station for it.

The ConvoyersConvoy AI team
Docs
ProtocolsIdentityProfilesSkillsPlaybooksRulesJudgmentsEventsAutomations

Use cases

One Product Station to handle any customer-sent agent request.

Each opens one product to every assistant, with only the actions you allow and the limits you set.

How it works

One layer between agents and everything they touch.

Convoy sits in the execution path. The rule is applied by the infrastructure, not suggested to the model, so it holds whatever the prompt says.

From

Your customers’ agents

  • ChatGPT
  • Claude
  • Gemini
  • Grok
  • Muse
  • OpenClaw
Through

Convoy Station

  1. AccessOne governed entry point per product or process, on every agentic protocol: MCP and API.
  2. DiscoverabilityOnly the tools that fit the current step are shown to the agent.
  3. ReliabilityEach action validated against your rules before it runs. Off-path calls are rejected and redirected.
  4. ContinuityContext carried across channels, tools and sessions.
  5. TraceabilityEvery transaction logged, with what was asked, what ran, why and by whom.
To

Your product

  • Store
  • Catalogue
  • Orders
  • Booking
  • Payments
  • Accounts
03 · Expose

Open it on your customers’ AI channels.

The ChatGPT and Claude directories, AI embedded in your site, your own AI app.

ChatGPTClaudeEmbedded AI
02 · Build

We build your Convoy Station.

Skills and playbooks from your APIs, your rules as policies, tested before it goes live. It speaks every agentic protocol: MCP and API.

skillsplaybookspoliciesMCPAPI
01 · Connect

Point us to your product’s APIs.

OpenAPI, GraphQL or plain REST. Nothing to rewrite.

OpenAPIRESTGraphQL

What we measure

Agents served, nothing broken.

Every station is watched on our four axes, call by call: accessibility, discoverability, reliability and context continuity. The charts show what each measure looks at.

Accessibilityof agent requests served on their protocol
Discoverability↓ tokensonly the tools the current step needs
Reliability0request errors from agents
Context continuity↑ resumedjourneys picked up across steps, sessions and channels, not restarted

Use cases

One Process Station per workflow, ready for any agent.

Each links the tools a process needs and holds every agent to its steps, its limits and its approvals.

How it works

One layer between agents and everything they touch.

Convoy sits in the execution path. The rule is applied by the infrastructure, not suggested to the model, so it holds whatever the prompt says.

From

Your teams’ agents

  • Claude
  • ChatGPT
  • Cursor
  • LangChain
  • Grok
  • OpenClaw
Through

Convoy Station

  1. AccessOne governed entry point per product or process, on every agentic protocol: MCP and API.
  2. DiscoverabilityOnly the tools that fit the current step are shown to the agent.
  3. ReliabilityEach action validated against your rules before it runs. Off-path calls are rejected and redirected.
  4. ContinuityContext carried across channels, tools and sessions.
  5. TraceabilityEvery transaction logged, with what was asked, what ran, why and by whom.
To

Your systems

  • CRM
  • ERP
  • Databases
  • Slack
  • Jira
  • Drive & docs
03 · Expose

Open it on your teams’ AI channels.

Claude, ChatGPT, Slack and your internal copilot, behind your SSO.

ClaudeChatGPTSlackCopilot
02 · Build

We build your Convoy Station.

Skills and playbooks per role, approvals and limits as policies, every call traced. It speaks every agentic protocol: MCP and API.

skillsplaybookspoliciesMCPAPI
01 · Connect

Point us to your systems’ APIs.

CRM, ERP, databases and internal tools, with the roles they already use.

RESTGraphQLSSO roles

What we measure

Agents at work, inside your rules.

Every station is watched on our four axes, call by call: accessibility, discoverability, reliability and context continuity. The charts show what each measure looks at.

Accessibilityof agent calls through one governed entry
Discoverability↓ tokensonly the tools each task needs
Reliability0actions outside your policies, all stopped
Context continuity↑ resumedthreads picked up at the step they stopped

Get started

Agents are already using your tools.
Decide how.

Set up your first station in days, on the APIs you already run. Every agent that comes in plays by your rules.

Build your first stationSee Stations Gallery
Or start with a free governance audit →