Iterations
0of 25
025 · stop
A hard cap on tool-call turns. The agent can’t reason itself into a loop you didn’t allow.
budget(turns=25)Agents
Agents inside workflows. Workflows inside agents.
Let Claude, ChatGPT or your own agents run the workflows you’ve built, as tools with your rules already inside. Four limits the engine enforces, and approvals that can wait for days.
Lyhnis speaks MCP on both sides. Expose authored workflows as typed tools to any MCP client; equip your agents with external tools from any MCP server. Four engine-enforced budgets and HITL gates that pause durably.
MCP both directionsWorkflows as toolsFour budgets engine-enforced
Scroll to see it agent.runclaude-sonnet · refund deskTurn 0 / 25
Tools you declared: verify_identity, process_refund, ledger_mcpSimulation · illustrative
Budgets, checked before every callReady
Iterations0 / 25
Tokens0 / 50,000
Wall-clock0 / 05:00
Cost0 / $0.500
Rollup covers every child call$0.000
01The boundary
Sub-workflows as tools.
The agent never gets raw access to your systems. It gets the workflows you wrote, with checks, approvals and the audit trail already inside. The workflow is the safety rail.
The agent’s tool vocabulary is author-explicit: workflows with validation, HITL gates and audit-by-construction wired in before the agent sees them.
The agent
Refund desk
claude-sonnet · 8 turnsDecides which tool to call and in what order. Nothing else.
Tools you declared
verify_identityvalidationauditprocess_refundvalidationapprovalauditledger_mcp.postcredentialedaudit
Your systems
- Stripe
- Ledger
What the agent can’t do
- Reach a raw API key or credential
- Skip an approval gate inside a tool
- Invent a tool you didn’t declare
- Exceed a budget by reasoning about it
- Write to the audit trail, or around it
What the agent can do
- Call the workflows you listed as tools
- Chain them in an order you didn’t script
- Read the typed results and decide
- Call external MCP tools you credentialed
- Hand off to a person and wait
02MCP
Both directions of the bridge.
Outside agents can call your workflows in. Your agents can call outside tools out. Both land in the same audit trail.
One protocol, two roles: a per-tenant MCP server for inbound calls, MCP credentials for outbound ones, one audit ledger for both.
Inbound · expose workflowsLive
ClaudeChatGPTAny MCP client
# your company’s MCP endpoint (illustrative) https://mcp.lyhnis.com/t/meridian # JWT-bridge auth, scoped per client Authorization: Bearer eyJhbGciOi… # tools the client sees process_refund workflow verify_identity workflow lookup_invoice workflow
Per-company serverhard isolation
JWT-bridge authscoped tokens
Every call auditedactor=agent
Your company on LyhnisOne audit trail, both directions
Outbound · call external toolsLive
Linear MCPYour MCP server
# a credential for an external MCP server { "name": "linear_mcp", "type": "MCP_OAUTH", "server_url": "https://mcp.linear.app" } # the agent step that may use it { "activity": "agent.run", "tools": ["process_refund", "mcp:linear_mcp"], "budget": { "iterations": 10, "tokens": 50000 } }
EncryptedFernet, per company
Password to rotatestep-up, always
Every call loggedserver + tool
03Budgets
Four limits the agent can’t talk past.
The limits live in the engine, not in the prompt. Set them once on the outer agent and they cover every agent and workflow it starts.
Budgets are enforced by the engine before every model call. Set them on the outer agent.run; the parent covers every child it spawns.
Ready
Tokens
0of 50,000
050k · stop
Input and output across the whole run, counted through every nested agent.
budget(tokens=50_000)Wall-clock
0of 05:00
05m · stop
Only active reasoning time counts. Waiting for a person costs nothing.
budget(wall_clock="5m")Cost
0of $0.500
$0$0.50 · stop
Real money: nested agent spend, model fees and metered tool use.
budget(cost_usd=0.50)Recursive rollup.
A parent’s budget includes every child agent, every sub-workflow and every external tool call. Cap the whole tree from the top, and auditors get one number per run.
See the v2.35 rollupagent.run · refund deskparent · budget set here$0.021verify_identitysub-workflow$0.000agent.run · assess photoschild agent$0.014ledger_mcp.postexternal MCP tool$0.007
04One run
Two kinds of actor. One audit trail.
An agent-started refund from start to finish: called over MCP, paused two days for a person, then rolled up to one cost. Every step records who acted and how.
Run run_01JX9 · refund deskPick a step to see its recordIllustrative run
Agent
Person
System
14:2214:23+2d 04:12+2d 04:13
Actor typing
Every entry is an agent, a person or the system. Never an anonymous service account.
Resolution source
Which channel answered, which model decided and which credential was used.
Cost contribution
Spend per step, rolled up to the run. No rebuilding it from logs.
05Status
What’s live on this page.
Every agent capability above carries its label. If it isn’t live, it says so.
7of 10 capabilities live
Checked against the product in September 2026
- agent.run inside any workflowAn agent is one step, with the tools and limits you set.Live
- Sub-workflows as toolsThe agent calls workflows you wrote, never a raw API.Live
- MCP server for your workflowsClaude, ChatGPT or any MCP client can call them as typed tools.Live
- MCP tools from outsideAgents call external MCP servers you have credentialed.Live
- Four budgets the engine enforcesIterations, tokens, wall-clock and cost.Live
- Recursive cost rollupOne number per run, covering every child.Live
- Approval gates inside agent toolsThe agent waits; it can’t skip the gate.Live
- Declaring tools in PythonDecorators for the tool vocabulary.To confirm
- Approve buttons inside Slack or TeamsDecisions happen on the email link.Not offered
- Streaming the audit trail to a SIEMExport to CSV instead.Not offered
30-minute call with an engineer
Build the next agent
worth shipping.
Bring the agent you want to trust with money. We’ll wire its tools and budgets with you.