Agents can't skip ahead
Define a task as a DAG of steps with explicit dependsOn ordering. The system enforces the order, so agents can't skip ahead or work out of turn.
Early access
One source of truth for every AI agent.
Claude Code, Codex, and any MCP agent connect to one shared task graph — steered through plan → execute → verify with human approval gates. Meshive is the single source of truth for what's planned, done, and proven.
See how it works No spam, just launch news.
What Meshive does
A workflow execution engine for agents
Not a tracker built for humans. Not raw LLM logs. Meshive defines how work should happen, agents follow it, and the system proves every step was done.
You hold the gates
Block execution before risky steps. The agent calls request_gate and waits while you approve or reject in the Inbox — humans decide the hard calls.
Any agent plugs in over MCP
Claude Code and other agents connect with a per-agent bearer token and drive work through a typed MCP tool surface. Ordering is enforced at the tool layer.
Proof every step was done
Every step lands an artifact — plan, diff, test report, verification. Watch status flow across the node graph in real time on the operator desk.
How it works
The agent loop, enforced
Agents connect over MCP and drive the workflow through a typed tool surface. You stay in command at the gates.
- 01
Design the task
An operator (or an agent) lays out the plan → exec → verify DAG. Each step declares what it depends on.
create_task()get_task() - 02
Agents execute in order
The agent asks what's ready, reads the step's contract, does the work, and attaches an artifact. The tool layer refuses out-of-order steps.
next_ready_steps()get_step_brief()start_step()complete_step() - 03
You hold the gates
At a gate the agent pauses and waits. You approve or reject in the Inbox; on approval the graph advances toward Done.
request_gate()get_gate_status()
For operators
Steer the work. Prove it was done.
Built for the engineer or lead overseeing agent-driven work in a codebase: define how a feature should be built, hand it to an agent, and approve the decisions that count.
Join the waitlist- Not a tracker for humans It assumes agents do the work — tasks are graphs of agentic steps, not tickets in a column.
- Not raw LLM logs You see milestones, artifacts, and gates — not a firehose of individual tool calls.
- Not a fire-and-forget runner Human gates pause execution at the moments that matter, so judgment stays in the loop.
FAQ
Questions
What is MCP?
The Model Context Protocol is an open standard for connecting AI agents to tools. Meshive exposes its workflow as MCP tools, so agents like Claude Code can drive tasks directly — discovering ready steps, reading briefs, completing work, and requesting gates.
Which agents and models work with Meshive?
Any MCP-capable agent. Each agent connects with its own bearer token, so actions are attributed to the connection that made them. Claude Code is the primary target during early access.
Do agents define the workflow, or do I?
Either. Tasks are a DAG of steps with explicit dependencies — you can lay them out yourself, or have an agent propose the graph with create_task. The system enforces the ordering regardless of who designed it.
When does it launch?
Meshive is in active development. Join the waitlist and we'll reach out as early-access spots open up.
Early access
Be first to steer your agents.
Meshive is in active development. Join the waitlist and we'll reach out as spots open up.