Zenith MCP server
Commit Sep 6, 2026Turns coding agents into multi-agent mission orchestrators to manage long-running tasks through adaptive planning and verification.
Also known as Intelligent-Internet/zenith
- Stars
- 403
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Sep 6, 2026
- License
- Apache-2.0
What is Zenith MCP?
Zenith turns coding agents into multi-agent mission orchestrators to manage long-running tasks through adaptive planning and verification.
Summary generated from the project's README
About Zenith MCP server
Developers managing complex, long-running coding tasks can use Zenith to transform standard coding agents into multi-agent mission orchestrators. This harness prevents premature completion by using an orchestrator session that reads task state to decide whether to spawn workers, deploy testers, register reusable skills, or replan. It is designed for work that may span days or weeks, specifically addressing the failure mode where agents stop before a task is truly finished.
To use Zenith, you must have Python 3.11+, uv, and Node.js 22+ installed. It integrates with existing agent environments like Claude Code, Codex, or Hermes. After initializing a workspace with the uv run zenith init command, you can run missions through your preferred agent. The system utilizes ACP adapters, such as @agentclientprotocol/claude-agent-acp, to manage worker and validator contexts. This approach provides a more adaptive orchestration method compared to the RALPH baseline, achieving high performance with lower per-task costs.
What you can do with it
- 1Preventing premature task completionUse the orchestrator to ensure coding agents do not stop working before a long-horizon requirement is fully met.
- 2Automated multi-agent orchestrationDeploy a single orchestrator to dynamically manage sub-agents for specialized roles like testing or execution.
- 3Adaptive mission planningHandle complex software engineering tasks that require constant replanning and verification based on current project state.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 14, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableintelligent-internet/zenith is public and reachablegithub.com/intelligent-internet/zenith ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)Apache-2.0spdx.org/licenses/Apache-2.0.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-09-06github.com/intelligent-internet/zenith ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubZenith: A Continuous-Improvement Harness for Long-Running Tasks
Zenith is an agent harness for work that may run for days or weeks, where the dominant failure mode is premature completion rather than inability to make progress. It runs a coding agent (Claude Code, Codex, or Hermes) as a multi-agent orchestrator over MCP/ACP: one orchestrator session reads task state each turn and decides whether to spawn workers and testers, register reusable skills, replan, or stop.
This repository contains the Zenith harness (zenith/) and the Intelligent Internet technical report (2026) behind it.
Abstract
Long-running agents often fail not because they cannot make progress, but because they stop before the task is truly complete. We tested five harness designs across eight long-horizon tasks to isolate the control mechanisms that matter: repeated gap-finding, revisable planning, independent verification, adaptive orchestration, and stopping discipline.
RALPH is the strongest simple baseline because it forces each new session to reopen the gap between the current project state and the original requirement. But RALPH is expensive and has no principled stopping rule.
Our Zenith method keeps the useful parts of repeated review while making the loop adaptive: the orchestrator dynamically allocates workers, testers, reusable skills, replanning, and stopping decisions. In this study, Zenith achieved the best mean rank while using less than half of RALPH's per-task cost.
Quick Start
Option 1 — Let your agent install it
Copy this prompt into Claude Code or Codex:
/goal Read the readme at https://github.com/Intelligent-Internet/zenith, detect if using Claude Code, Codex, or both, install requirements, install and run Zenith (i.e. uv run zenith, as in the readme), and create a new skill called /zenith — when used (along with an additional prompt) it will call the skill: the minimum skill content should be: """First read .claude/orchestrator_prompt.md and treat it as your primary role, then use Zenith to run this mission.""" Afterwards, you can add information about the Zenith harness, based on the readme and the technical report (inside the repo), and info on how to start Zenith if it's not already running. Change the skill to use .codex when using it in Codex. If both harnesses are available, make sure to add the skill to both of them correctly. In fact, there might be more harness options (Hermes, for example). See what is supported in zenith/src/zenith_harness/providers.py, and for those that you detect are present, add their skills correctly. When finished, confirm to me that Zenith is installed, running, and ready, explain a bit about Zenith, and why and when to use it.
This will:
- install Zenith with its requirements
- start Zenith using
uv - create a
/zenithskill for each agent harness it detects
Then, in Claude Code or Codex, type:
/zenith <your instruction or query>
Option 2 — Install manually
Requirements
- Python 3.11+
uv- Node.js 22+ and
npm - Claude Code, Codex, or Hermes (see
providers.pyfor the supported set)
Install
cd zenith
uv sync
uv run zenith --help
Install the ACP adapters globally for the agents you want Zenith to run:
# Claude workers/validators
npm install -g @agentclientprotocol/claude-agent-acp
command -v claude-agent-acp
# Codex workers/validators
npm install -g @agentclientprotocol/codex-acp
command -v codex-acp
Initialize a workspace
Initialize the project workspace Zenith should operate on. This is your target app/repo, not the Zenith source checkout:
# Claude Code, from this Zenith checkout
uv run zenith init --workspace-dir /path/to/your-app --agent claude
# Or Codex, from this Zenith checkout
uv run zenith init --workspace-dir /path/to/your-app --agent codex
Run a mission
Start your agent from the initialized project workspace:
cd /path/to/your-app
claude
# or
codex
Then ask the agent to read the generated orchestrator prompt (use .codex/orchestrator_prompt.md for Codex):
First read .claude/orchestrator_prompt.md and treat it as your primary role, then use Zenith to run this mission.
<your instruction or query>
How Zenith Works
A single orchestrator session reads task state each turn and decides what to do next: spawn worker or tester subagents, register a reusable skill, replan, or stop. Workers and testers run in their own contexts and report back; the orchestrator integrates their results before the next decision.
Results
Frontier SWE Benchmark
On the Frontier SWE benchmark, Zenith — running on GPT-5.5 — ranks first overall, leading on implementation, performance, and dominance against frontier models paired with their native harnesses.
| # | Model | Harness | Avg rank ↓ | Dominance ↑ | Implementation ↓ | Performance ↓ | Research ↓ |
|---|---|---|---|---|---|---|---|
| 1 | GPT-5.5 | Zenith | 2.06 | 92% | 1.60 | 1.89 | 3.33 |
| 2 | Claude Fable | Claude Code | 2.71 | 88% | 1.80 | 2.11 | 6.00 |
| 3 | Claude Opus 4.8 | Claude Code | 5.06 | 71% | 4.20 | 5.56 | 5.00 |
| 4 | GLM-5.2 | Claude Code | 5.31 | 69% | 5.60 | 6.50 | 1.67 |
| 5 | GPT-5.5 | Codex | 5.53 | 68% | 7.40 | 4.44 | 5.67 |
| 6 | Claude Opus 4.7 | Claude Code | 6.35 | 59% | 5.00 | 7.00 | 6.67 |
| 7 | Claude Opus 4.6 | Claude Code | 7.53 | 52% | 7.60 | 7.56 | 7.33 |
| 8 | GPT-5.4 | Codex | 8.06 | 50% | 7.20 | 9.67 | 4.67 |
| 9 | Composer 2.5 | Cursor CLI | 9.35 | 38% | 7.80 | 11.11 | 6.67 |
| 10 | Gemini 3.1 Pro | Gemini CLI | 9.65 | 37% | 11.80 | 7.44 | 12.67 |
| 11 | GLM-5.1 | Claude Code | 10.88 | 29% | 10.80 | 11.00 | 10.67 |
| 12 | DeepSeek V4 Pro | Claude Code | 11.00 | 27% | 10.80 | 11.11 | 11.00 |
| 13 | Kimi K2.5 | Kimi CLI | 11.65 | 24% | 13.00 | 10.22 | 13.67 |
| 14 | Kimi K2.6 | Kimi CLI | 11.82 | 25% | 10.40 | 12.78 | 11.33 |
| 15 | Qwen3.6-Plus | Qwen Code | 12.47 | 21% | 15.00 | 10.67 | 13.67 |
Metrics as reported by the Frontier SWE leaderboard. Rank columns: lower is better. Dominance: higher is better.
Ablation Study
To isolate the control mechanisms that matter, we compared Zenith against RALPH and three reduced harness variants across eight long-horizon tasks. Zenith achieves the best mean rank at less than half of RALPH's per-task cost.
| Method | Mean rank ↓ | Mean cost (USD/task) ↓ | Wins (of 8) ↑ |
|---|---|---|---|
| One-session | 5.00 | $22.21 | 0 |
| Plan-RALPH | 4.00 | $161.53 | 0 |
| Milestone-RALPH | 2.88 | $209.47 | 0 |
| RALPH | 1.75 | $407.58 | 3 |
| Zenith | 1.38 | $175.68 | 5 |
A "win" is a task on which the method ranked first; the eight wins partition the eight benchmark tasks.
Repository Layout
| Path | Contents |
|---|---|
zenith/ |
The Zenith harness — Python package (zenith-harness), CLI, MCP server, bundled prompts and skills, tests |
technical_report/ |
From RALPH to Zenith technical report — PDF, LaTeX source, and figures |
Contributing
Contributions are welcome — see CONTRIBUTING.md for the development setup and PR guidelines, and SECURITY.md for how to report vulnerabilities.
Citation
@techreport{ii2026zenith,
title = {From RALPH to Zenith: Designing Harnesses for Long-Running Agents},
author = {{Intelligent Internet}},
institution = {Intelligent Internet},
year = {2026},
type = {Technical Report},
url = {https://github.com/Intelligent-Internet/zenith}
}
License
The Zenith code is licensed under the Apache License 2.0. The technical report and its figures are licensed under CC BY 4.0.
Zenith MCP questions
What is Zenith?
Zenith is an agent harness designed for work that may run for days or weeks, preventing premature completion of long-horizon tasks.
Which agents does Zenith support?
Zenith supports Claude Code, Codex, and Hermes.
How does the orchestration work?
An orchestrator session reads the task state each turn to decide whether to spawn workers and testers, register skills, replan, or stop.