Provides tools for interacting with a Git repository, including tracing parent ancestry for candidate commit SHAs and inspecting Git history for provenance validation.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Compounding MCP initialize a swarm for tuning hyperparameters"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
A model-agnostic MCP server for persistent agent contributions and numerical experiments. It uses Boruna workflows for capability-gated checks and evidence bundles; Node.js provides the MCP transport and Git-history inspection.
Experimental proof of concept. A workflow operation is validated by a Boruna
.axstep. The wrapper then records JSON state under.boruna/swarm/<id>/. Version 0.1 supports one active server process per working directory and does not create or merge Git commits.
Node.js 20+
Boruna 4.2+
Git repository as working directory
Related MCP server: tower-mcp
Download the compounding-mcp-0.1.0.tgz release asset, then run:
npm install --global ./compounding-mcp-0.1.0.tgzSet up any MCP client with:
{
"mcpServers": {
"compounding": {
"command": "compounding-mcp",
"env": { "COMPOUNDING_WORKDIR": "/path/to/repository" }
}
}
}If Boruna is not on PATH, set BORUNA_BIN to its executable path.
swarm_init: create a task spec and episode budget.
swarm_status: read budget, workers, contributions, nominations, and best score.
swarm_announce: publish worker scope, deliverable, success condition, and abort condition.
swarm_commit: validate provenance fields with Boruna, then record the contribution, score, and validation note.
swarm_nominate: record a candidate vote and rationale.
swarm_abstain: record why a worker has no useful work; does not consume budget.
swarm_ancestry: trace Git parent ancestry for a candidate SHA (requires a committed candidate in the selected repository).
All records live in .boruna/swarm/<swarm_id>/. The Boruna policy grants only fs.read and fs.write within .boruna/swarm.
The MCP server is model-agnostic, but each client must launch it through its MCP configuration.
swarm_commit stores provenance metadata; it does not apply candidate code or create a Git commit.
The server trusts caller-supplied benchmark scores and validation descriptions. It does not run the numeric benchmark.
informed_by validation checks SHA syntax and requires the physical parent to appear in the list. Full cross-branch graph analysis is planned.
Keep only one active server process per working directory. The Boruna workflow interface has fixed staged input paths.
Nomination counts are coordination metadata, not correctness evidence.
npm ci
npm test
npm pack --dry-runCheck Boruna assets with:
for d in workflows/swarm_*; do boruna workflow validate "$d"; done
find workflows -name '*.ax' -print0 | xargs -0 -n1 boruna lang check --jsonPush a version tag such as v0.1.0. GitHub Actions runs tests, validates workflows, builds the npm tarball, and attaches it to a GitHub Release. No GitHub login is needed by end users to download a public release.
Release a worker slot without spending an episode, with a non-empty reason.
Inputs: reason (required), worker (required), swarm_id (required).
Trace Git parents and informed_by dependencies from a candidate SHA in the configured repository.
Inputs: swarm_id (required), candidate_sha (required).
Publish worker scope, deliverable, success condition, and abort condition before editing.
Inputs: gap (required), abort (required), worker (required), success (required), +2 more.
Record a validated contribution with parent and semantic informed_by SHAs.
Inputs: score (required), parent (required), worker (required), message (required), +3 more.
Create a persistent swarm with a task spec and episode budget.
Inputs: task (required), budget (required), metric (required), swarm_id (required), +1 more.
Nominate an existing candidate with a rationale. This records a vote, not a quality proof.
Inputs: worker (required), swarm_id (required), rationale (required), candidate_sha (required).
Read state, in-flight workers, nominations, and best validated score. Votes are not correctness evidence.
Inputs: swarm_id (required).
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 7 tools
Each tool targets a clearly distinct action within the swarm lifecycle: init/status manage the swarm, announce/commit/nominate/abstain handle worker contributions, and ancestry traces lineage. The descriptions explicitly distinguish similar concepts (e.g., nominate is a vote, not a quality proof; commit records a validated contribution). No overlapping purposes.
All tool names follow the same `swarm_` prefix followed by a concise snake_case verb (abstain, ancestry, init, status, announce, commit, nominate). The convention is uniform with no deviations or mixed styles.
Seven tools is well within the ideal 3-15 range and each maps to a distinct coordination operation in the swarm protocol. No tool feels redundant or out of scope for the server's stated purpose.
The surface covers the core swarm lifecycle: create a swarm, announce work, commit validated contributions, nominate candidates, abstain, trace ancestry, and read status. Minor gaps remain, such as no explicit tool to list/update/close swarms or fetch a specific candidate's full details, but these are workaroundable via status and ancestry.
Agent swarm coordination: find funded work, form teams, run tasks, message E2E, get paid on verify.
Source control for coding agents: declare an intent, settle overlaps before code, verify each push.
Repository evidence for agents before they adopt dependencies, enter codebases, compare, or merge.
Machine-native research commons for agent evidence, discovery, rooms, and bounded research quests.