SkillAgentSearch skills...

Panter-ai

One leader. Every model. Every MCP. Open-source MCP control plane that turns Claude, ChatGPT, Grok and Gemini into one engineering team: durable delegation, scoped GitHub grants and an independent review gate on the exact commit. No model APIs.

Install / Use

claude mcp add riquelmechile -- npx -y github:riquelmechile/Panter-ai

If the server publishes to npm under a different name, use that package instead — check the repo README.

About this skill
🔌

MCP Server

Model Context Protocol server

Quality Score

81/100

Category

Automation

Supported Platforms

Claude Code
Claude Desktop
Gemini CLI

Our assessment of Panter-ai

Panter-ai scores 81/100 on our quality scale, 2333rd of 2,775 Automation skills we index.

Its MCP Server is 16 KB long, well organised into 16 sections with 2 code examples: a thorough specification that gives an agent plenty to work with.

It has 3 GitHub stars, so there is little community track record yet; judge it on its content.

Substance
30/30
Structure
18/20
Description
15/15
Adoption
3/20
Freshness
15/15

Maintenance, license and trust

  • The repository was last updated 3 days ago, so Panter-ai is actively maintained.
  • It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
  • Its trust signals score 92/100, with 1 caution from licensing, adoption, age or documentation. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.

Safety scan

No issues found

Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful.

AI review by kimi-k2.7-code on 2026-11-16. Automated pattern scan on 2026-11-16. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.

Panter-ai compared with similar skills

All 4 of these similar skills score higher than Panter-ai; compare them before choosing.

SkillScoreStarsUpdatedFormat
Panter-ai (this skill)by riquelmechile8133d agoMCP Server
Agent-Reachby Panniantong100112.4k9d agoCLAUDE.md
rufloby ruvnet10081.9ktodayCLAUDE.md
headroomby headroomlabs-ai10079.9ktodayCLAUDE.md
CowAgentby zhayujie10048.0ktodayCLAUDE.md

Frequently asked questions

How do I install Panter-ai?
Run claude mcp add riquelmechile -- npx -y github:riquelmechile/Panter-ai. The install tabs above show the steps for each supported agent.
Which AI agents does Panter-ai work with?
It is written for Claude Code, Claude Desktop and Gemini CLI, as a MCP Server file. Other agents that read the same format can often use it too.
Is Panter-ai safe to use?
Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. An AI review of the same text found nothing harmful. It is MIT-licensed and scores 92/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
Is Panter-ai still maintained?
The repository was last updated 3 days ago, so Panter-ai is actively maintained.
<p align="center"> <img src="docs/assets/panter-hero.svg" alt="Panter-ai: one leader, every model, every MCP. Animated banner with Panter, a black panther with a mouse face whose tail ends in a glowing connector." width="100%"> </p> <p align="center"> <img src="docs/assets/panter-typing.svg" alt="A terminal types a task moving through Panter-ai: Claude delegates, Panter wakes Grok, Grok claims the work and opens a pull request with evidence, the reviewer approves the exact commit, Claude merges." width="100%"> </p> <p align="center"> <b>English</b> · <a href="README.es.md">Español</a> </p> <p align="center"> <a href="LICENSE">MIT</a> · <a href="CHANGELOG.md">Changelog</a> · <a href="docs/SELF_HOSTING.md">Self-hosting</a> · <a href="docs/OWNER_MODE.md">Owner mode</a> · <a href="docs/CANONICAL_HARNESS_POLICY.md">Policy</a> · <a href="docs/DOWNSTREAM_MCP.md">Downstream MCPs</a> </p>

Panter-ai

Panter-ai is an open-source MCP server that turns several AI apps into one engineering team. One app leads: it plans and hands out the work. The others take that work (exploring, implementing, verifying, reviewing) and send back results. Panter keeps the plan, the permissions and the evidence in one durable ledger, so nobody has to take a model's word that something was done.

Panter never calls a model API. Each app (Claude, ChatGPT, Grok, Gemini) connects with its own account and runs its own model. The server holds no model API keys and runs no inference: it coordinates, records and gates.

Who it is for: anyone who already uses several AI apps and wants them to work on the same projects with clear roles, real GitHub pull requests and an independent review before anything is merged.

Meet Panter

The mascot is a panther with a mouse's face. The panther is patient and precise and does not let go of a job until it is done. The mouse is small and curious and gets into every corner. Its tail ends in a connector node, because that is how Panter works: every app and every MCP server plugs into it.

<p align="center"> <img src="docs/assets/panter-mascot.svg" alt="Panter, a black panther with mouse ears, a pink nose and whiskers, and a tail ending in a teal connector" width="180"> </p>

How it fits together

<p align="center"> <img src="docs/assets/panter-architecture.svg" alt="AI hosts connect to Panter-ai over MCP with OAuth. Panter runs the harness, review gate, mesh and workforce, GitHub workforce, memory and tool surface, reaches downstream MCP servers through a gateway, and gates merges through GitHub." width="100%"> </p>
  • AI hosts (the apps) connect to https://<your-domain>/mcp with OAuth. You approve each one on a consent screen and bind it to a host name: claude, gpt-a, gpt-b, grok or gemini. PANTER_LEAD_HOST picks the leader; every other host is a worker.
  • Panter-ai runs the work. It picks the next phase, sends it to a host that can do it, records what came back and refuses to move on without evidence.
  • Downstream MCP servers (your own tools, each in its own repository and service) sit behind Panter. Hosts reach them through one gateway, with a token per server, write grants and an audit log.
  • GitHub is where work lands: pull requests, CI and an independent reviewer account approving the exact commit.

How a task flows

<p align="center"> <img src="docs/assets/panter-flow.svg" alt="Seven steps: the leader delegates, Panter queues the work with a scoped GitHub grant, wakes the worker, the worker claims it with a lease, works in its own app on its own branch and opens a pull request, completes with evidence, and the pull request goes through review and merge." width="100%"> </p>
  1. The leader delegates. It calls panter_harness_tick (the automatic harness picks the phase) or delegates directly with panter_mesh_delegate.
  2. Panter queues it as a durable work item and, if the work touches GitHub, attaches a permission scoped to that one item and one repository.
  3. Panter wakes the worker through the route its app supports (next section).
  4. The worker claims it with a fenced lease and heartbeats, so only one attempt ever owns the work.
  5. The worker does the work in its own app. On GitHub it commits to a new branch, panter/<host>-<id>, and opens a pull request.
  6. It completes with evidence. Panter stores the result, sends the leader a receipt and revokes the GitHub permission.
  7. Review and merge. CI and an independent reviewer check the exact commit; the leader merges.

If a worker does not pick the work up in time (3 minutes by default, 20 for mail), a harness phase falls back to the leader, which does it natively. A sleeping worker never blocks a run.

How workers wake up

<p align="center"> <img src="docs/assets/panter-wake.svg" alt="Panter rings each worker through its own route: ChatGPT hosts through an MCP event subscription and a ChatGPT Work automation, Grok through a bearer webhook, Claude through a Claude Code Routine, Gemini through a Gmail relay or a scheduled task, and any host with a chat open through a live MCP listener." width="100%"> </p>

| Host | How Panter wakes it | Setup | |---|---|---| | GPT-A, GPT-B (ChatGPT) | MCP event subscription that triggers a ChatGPT Work automation | CHATGPT_NATIVE_MCP_EVENTS.md | | Grok | Bearer webhook to its Grok Bot Routine | GROK_BOT_WAKE.md | | Claude | Claude Code Routine | CLAUDE_ROUTINE_WAKE.md | | Gemini | Gmail relay or a scheduled task (never writers or sensitive missions) | GEMINI_MAIL_WAKE.md | | Any host with a chat open | Live MCP listener (long poll) | WORKFORCE_WAKE.md |

A ring the host accepted is not proof of work. Panter only trusts the worker's claim and its result. panter_workforce_status shows each worker's wake routes and what it finished in the last 24 hours.

Reviews and the delivery gate

<p align="center"> <img src="docs/assets/panter-delivery-gate.svg" alt="A pull request goes through the delivery gate. Always blocking: exact commit, open non-draft PR on the authorized base, required CI and the independent reviewer approving that commit. Blocking by default, warnings in owner mode: branch naming, type label, approved issue, commit format, PR size. Then the leader merges. Workers can never merge, approve, force-push, write protected branches, CI or git config, or read secrets." width="100%"> </p>
  • 4R review: four lenses (risk, readability, reliability, resilience), run natively or on the hosts bound to them.
  • Judgment Day (on request): two blind judges, never on a host that wrote the candidate.
  • Panter-RDD: acknowledges a review only when it is bound to the exact commit, its CI and attested transports.
  • Independent reviewer account: a second GitHub account, set in PANTER_GITHUB_INDEPENDENT_REVIEWER, gives the formal approval on the exact head commit. See INDEPENDENT_REVIEW_IDENTITY.md.

A review never grants permission to merge or deploy. Those stay with the leader and its operator.

Owner mode

If you operate every connected app yourself, set PANTER_OWNER_MODE=1 (OWNER_MODE.md):

  • the leader delegates writers directly, without an operator write grant, and a writer may cover the whole repository;
  • wake routes accept writers, and work queues behind a busy worker instead of being refused;
  • panter_mesh_delegate_auto picks the least-loaded reachable worker;
  • keyword-based risk becomes advice;
  • the delivery gate turns process conventions into warnings and keeps blocking on the exact commit, the PR, the base, CI and the independent reviewer (which owner mode requires).

The worker limits in the diagram never change.

Your repository and your server

<p align="center"> <img src="docs/assets/panter-repo-vs-server.svg" alt="The public repository holds code, tests, docs and example configs with placeholders. Your server holds panter.env, downstream.json, the data directory and backups. deploy.sh installs an exact commit on the server; nothing flows back." width="100%"> </p>

This repository is generic: no accounts, domains, tokens or passwords of any deployment. Everything that is yours lives on your server:

| File | What it holds | |---|---| | panter.env | secrets, accounts, domain, leader, owner mode, wake routes | | downstream.json | the MCP servers you connect behind Panter | | data directory | SQLite ledger, OAuth clients and tokens, audit | | backups | taken before every deploy and every day |

To connect a new MCP server, add it to downstream.json, put its token in panter.env and restart. The repository does not change. A CI hygiene test fails if a tracked file carries a real credential or an example config carries a real value. Step-by-step setup: SELF_HOSTING.md.

What it does

| Area | What you get | |---|---| | Automatic harness | panter_harness_start and panter_harness_tick run ODD phases (discover, design, implement, verify, review) chosen by risk. The user never picks agents or skills. panter_harness_close stops a run for good. | | Review gate | 4R with four lenses, optional Judgment Day with two blind judges, and Panter-RDD bound to the exact commit, its CI and attested transports. | | Mesh and workforce | Durable delegation, fenced claims with leases, heartbeats, results, wake routes, pickup timeouts with native fallback, and panter_mesh_withdraw for queued work nobody started. | | GitHub workforce | Write access scoped to one work item and one repository: read, commit to its own new branch, open a PR, comment, open an issue. Revoked when the work ends, with an audit trail. | | Memory | Optional Engram project memory, mirrored from the harness ledger. | | Tool surface | Each role sees a short list: the leader 50 tools and a worker 32, the 8 domain routers included. Every tool stays callable by name or through its router. See TOOL_SURFACE.md. | | Downstream gateway | Register an MCP server once; it appears as one router, panter_ds_<name>, with timeouts, a circuit breaker, write grants, a signed actor identity on every call, OAuth client credentials, sensitive servers that keep no content, and an audit log. See DOWNSTREAM_MCP.md. | | Operations | One-command deploy of an exact commit with automatic rollback, a backup with a restore check before every deploy and every day, and a health check every 5 minutes that restarts the service. See OPERATIONS.md. | | Server actions (off by default) | The leader hands a worker a single-use permit to deploy one exact commit. The worker gets no shell: Panter checks that the commit is merged, green and independently approved, runs the same deploy script, and reports back. See SERVER_ACTIONS.md. |

Ground rules

These are enforced in code, not just written down:

  • No model APIs. The server never runs inference and stores no model API keys.
  • Evidence over claims. A host saying "tests pass" is recorded as reported, not verified. Approval needs the exact commit, CI and an independent reviewer.
  • Authorization is separate. Writing code, merging and deploying each need their own authorization. Reviews never grant them.
  • Workers stay in their lane. They never merge, approve, force-push, write the default or protected branches, CI or git config files, or read secrets.
  • The host owns the model. Panter requests a model, effort and parallelism (src/core/policy.ts); it never claims what a host actually ran.

Quick start

Requires Node.js 22.16 or later.

npm ci
npm run build
npm test
export PANTER_TOKEN="$(node -e "process.stdout.write(require('node:crypto').rando

Truncated for display — read the full file on GitHub.

Related Skills

View on GitHub
GitHub Stars3
CategoryAutomation
Updated3d ago
Forks0

Languages

TypeScript

Trust signals

92/100

From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.

1 low
Panter-ai — MCP Server: Install & Safety Check | SkillAgent