H5i

Run many coding agents. Merge one auditable result.

Agent ensembles work because independent attempts beat isolated guesses. h5i runs several coding agents on the same task, each in its own sandbox, sealed so they can't copy one another. It lets them peer-review, then a neutral verifier replays every candidate, runs the tests itself, and merges the one that actually passes. The whole run (prompts, models, commands, logs, policies, messages, and the verdict) is versioned in your repo under refs/h5i/*.

> Two heads are better than one.

Isolated per agentno file, branch, or port clashes Auto peer-reviewcross-agent discussion Rich dashboarddiffs, reviews, results Lives in your Gitrefs/h5i/* · no SaaS

Who it's for: platform, security, and DevEx leads rolling out Claude Code and Codex who want to run teams of agents and keep review and audit defensible as agents write more of the diff.

---

1. Install

curl -fsSL https://raw.githubusercontent.com/h5i-dev/h5i/main/install.sh | sh

Or build from source:

cargo install --git https://github.com/h5i-dev/h5i h5i-core

---

2. 60-Second Flow

2.1. Setup

Initialize h5i and wire the Claude Code / Codex hooks:

h5i init
h5i hook setup --write --wrap-bash --team
git add .
git commit -m "update hooks"

2.2. Track Prompts and Contexts

Once the hooks are registered, h5i versions your human prompts and every agent context step (reads, writes, thinking) as Git objects, trimming noisy tool output along the way (for pytest, just the failures) to cut up to 95% of the tokens while keeping the raw output recoverable.

h5i recall context show   # replay the captured prompts and agent context steps

Share it with h5i share push, or post an AI-usage summary (prompt quality, AI/human commit ratio, secret leaks, prompt injection, and more) to the pull request with h5i share pr post (needs the gh CLI).

h5i share push      # push the h5i metadata (refs/h5i/*) to your teammates
h5i share pr post   # post the AI-usage summary to the pull request (needs `gh`)

2.3. Sandboxed Environment

h5i gives each agent a secure, sandboxed worktree. Let it run with permissions off inside the box, then review its diff before anything lands on your branch:

h5i env create claude-env --profile agent-claude
h5i env shell claude-env
box$ claude --dangerously-skip-permissions
box$ exit

h5i env diff claude-env      # review what the agent changed in the box
h5i env propose claude-env   # turn the box's work into a reviewable proposal
h5i env apply claude-env     # merge the reviewed changes onto your branch

2.4. Run an ensemble

Create a team

h5i team auto-create qsort-demo

##### This `auto-create` corresponds to making and registering two sandboxed environments:
#
# h5i env create qsort-demo-claude --profile agent-claude
# h5i env create qsort-demo-codex  --profile agent-codex
#
# h5i team create  qsort-demo --base HEAD
# h5i team add-env qsort-demo env/human/claude-env --runtime claude
# h5i team add-env qsort-demo env/human/codex-env  --runtime codex
# h5i team status  qsort-demo                                 # note the generated agent ids

Dispatch one task to every agent:

echo "Implement Quick Sort from scratch in Python." | h5i team dispatch qsort-demo

Launch every agent in its own sandboxed environment. Each agent automatically starts working on the dispatched task:

# Terminal 1: Claude, running inside its own h5i sandboxed env.
h5i env shell env/human/qsort-demo-claude -- claude "$(h5i team bootstrap)" # `--dangerously-skip-permissions`

# Terminal 2: Codex, running inside its own h5i sandboxed env.
h5i env shell env/human/qsort-demo-codex  -- codex  "$(h5i team bootstrap)" # `--sandbox danger-full-access`

Each agent peer-reviews, and revises inside its own implementation:

h5i team auto-peer-review qsort-demo                       # sync → freeze → mutual grant → instruct

Merge the best one:

h5i team apply --agent                           # id from `team status`

##### Alternatively, replay each candidate, run the tests, merge the winner:
#
# h5i team verify   qsort-demo --agent  -- pytest  # id from `team status`
# h5i team finalize qsort-demo                               # explainable verdict (gates + smallest diff)
# h5i team apply    qsort-demo                               # merge the winner, gated on the verdict

2.5. Web UI

Monitor the status:

h5i serve

---

3. What h5i is, and is not

> h5i is not a Git replacement, a hosted SaaS / dev-environment, or just a sandbox.

Why not a hosted sandbox?: The whole point is that the workspace and its evidence live in your repo (refs/h5i/*): pushable, fetchable, offline, and yours. Codespaces, Coder, and E2B give you an environment; h5i gives you an auditable one, versioned in Git with no service to depend on.

Why naive agent teams break: In ML, ensembles beat the best single model: diverse estimators cut variance and won a decade of competitions. The same shift is coming to coding agents. But spawn several agents on one repo with no coordination layer and you don't get an ensemble, you get a pileup:

| Failure mode | What happens | h5i's answer | |---|---|---| | Environment conflict | agents overwrite/destory each other's files | a confined worktree per agent | | Token explosion | every agent re-reads the repo and runs tools | compressed tool logs | | Review overload | humans can't inspect every prompt or command | reviewer-ready PR |

---

4. Documentation

• Official Website: project overview, Pitch Deck
• Tutorials: guided workflows · Blog: design notes, audits, case studies
• [MANUAL.md](MANUAL.md) / man h5i: full command reference
• [CONTRIBUTING.md](CONTRIBUTING.md): we welcomes contributions of any kind.

---

5. Acknowledgements

h5i's token-reduction filters build on prior art, both Apache-2.0:

• rtk: the declarative output-filter rule files and the engine that runs them are derived from rtk.
• headroom: the log line-folding technique (collapse near-identical lines into one with a count) is reimplemented from headroom.

See [NOTICE](NOTICE) and [assets/filters/NOTICE](assets/filters/NOTICE) for full attribution.

6. License

Apache-2.0. See [LICENSE](LICENSE).

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: h5i-dev
• Source: h5i-dev/h5i
• License: Apache-2.0
• Homepage: https://h5i.dev

Install and usage instructions live in the source repository linked above.