fluxgit-mcp-sidecar

Safety-first read-only Model Context Protocol (MCP) server for Git.

AI agents inspect. FluxGit keeps control.

A Rust MCP server that exposes 29 carefully designed tools for AI code agents to navigate Git repositories without ever being able to silently mutate them. Built as the bridge between MCP-compatible agents and the FluxGit desktop application.

Why this exists

AI coding agents are increasingly asked to navigate real repositories: explain branch state, summarize diffs, find lost commits, recommend safe next steps. To do this well, an agent needs Git context that is richer than git status and structured enough to reason over. To do this safely, an agent must never be able to silently mutate refs, force-push, discard work, or apply patches without a human approving the consequence.

Other MCP Git servers face a choice: stay strictly read-only (limited utility) or expose write tools directly (dangerous — agents hallucinate, prompts can be poisoned, mistakes are destructive). This sidecar chooses neither. Reads are unrestricted; writes go through a write-with-UI-handshake protocol that the FluxGit desktop app implements: the agent proposes, FluxGit shows the preview, the user approves in the app, FluxGit executes through its safety pipeline with restore points and audit.

Related MCP server: GitPilot MCP

What's exposed

22 read-only tools

7 write-with-UI-handshake tools

All 6 operations dispatch through the FluxGit gateway when configured (the original 5 as of 2026-05-28; plan added 2026-06-10). The sidecar POSTs the proposal to the gateway's handshake server, the FluxGit app renders an "🤖 Requested by AI agent" approval card per operation type, and the sidecar polls until the user approves, rejects, or the proposal expires. When the gateway is not reachable, the sidecar returns write_handshake_pending (code 10003) so the agent can recommend the user perform the action in FluxGit UI.

All 6 require a free-text reason so the user sees the agent's justification in the approval modal. All 6 reuse the same gateway state machine (pending → approved → completed, terminal states rejected | failed | expired) and the same Tauri bridge in the UI. When an approved merge/rebase/reset executes, the completion result includes the captured restore point (beforeCommit/afterCommit/canUndo) so the agent can tell the user the change is reversible from FluxGit's Safety Timeline.

Write protocol details

Every operation.preview.* call follows the same wire protocol. Example for operation.preview.merge:

1. Sidecar POSTs the proposal:

POST /v1/mcp/operation/preview/merge HTTP/1.1
Host: 127.0.0.1:59647
Content-Type: application/json

{
  "previewId": "1f3c5b9a-...-uuid",
  "agentId": "external-mcp-sidecar",
  "operationType": "merge",
  "repoPath": "/Users/dev/projects/checkout",
  "sourceRef": "feature/cart-redesign",
  "targetRef": "main",
  "reason": "Cart redesign work is complete; tests pass on the feature branch.",
  "strategy": "merge",
  "requestedAt": "2026-05-28T11:42:09.512Z"
}

2. Gateway responds 202 Accepted:

{ "previewId": "1f3c5b9a-...-uuid", "status": "pending", "expiresAt": "2026-05-28T11:47:09.512Z" }

3. Sidecar polls every 1s for up to 60s:

GET /v1/mcp/operation/status/1f3c5b9a-...-uuid HTTP/1.1
Host: 127.0.0.1:59647

4. Gateway returns terminal state once user acts:

{
  "previewId": "1f3c5b9a-...-uuid",
  "operationType": "merge",
  "status": "completed",
  "result": {
    "commitSha": "9a8b7c6d...",
    "restorePointId": "rp_2026_05_28_1142",
    "conflicts": []
  }
}

The sidecar returns the result to the agent as isError: false. Any terminal status other than completed (rejected, failed, expired) returns isError: true with the structured payload, so the agent can report the rejection reason cleanly without inventing an outcome.

The same pattern applies to all 6 operations. Only the request body fields and the result shape differ; the polling, state machine, and error semantics are shared. The full per-operation contract ships with the FluxGit desktop app and is summarized at fluxgit.com/features/mcp-agent-git.

Boundary: free shell vs FluxGit-powered

The sidecar speaks MCP without FluxGit installed. Standard Git inspection works (status, refs, history, reflog, diff.text, etc). The tools that require FluxGit return JSON-RPC error code 10001 with an upgradeHint pointing the agent at the install/configure flow.

Tier classification:

• Free shell — work with local git only: repo.brief, repo.scope, repo.status, repo.refs, repo.branchStack, repo.history, repo.reflog, commit.details, worktree.changes, worktree.list, submodule.status, diff.text, conflict.read.

• Hybrid — work locally with documented fallback, enriched by FluxGit: fleet.radar, diff.semantic, diff.semanticFallbacks, repo.conflictPreflight.

• FluxGit-required — return gateway_not_configured without FluxGit because synthesizing them from local refs alone would mislead the agent: safety.timeline, safety.eventDetails, flux.latestRestorePoint, flux.restorePoints, flux.restorePointDetails.

• Write handshake — route through FluxGit UI approval via the gateway handshake server (as of 2026-05-28); return write_handshake_pending (code 10003) only when the gateway is unreachable or polling times out: the 6 operation.preview.* tools above.

Quick start

One-line install (puts fluxgit-mcp-sidecar on your PATH):

cargo install --git https://github.com/fluxgit-hq/fluxgit-mcp-server fluxgit-mcp-sidecar

Or build from a clone:

# Build
cargo build --release

# Run as MCP server (stdin/stdout transport)
./target/release/fluxgit-mcp-sidecar

Connect any MCP-compatible agent

Paste the generic block below into any MCP host config. No client-specific install required.

{
  "mcpServers": {
    "fluxgit": {
      "command": "/absolute/path/to/fluxgit-mcp-sidecar",
      "env": {
        "FLUXGIT_GATEWAY_ADDR": "127.0.0.1:14660",
        "FLUXGIT_MCP_AUDIT_LOG": "/optional/path/to/audit.jsonl"
      }
    }
  }
}

FLUXGIT_GATEWAY_ADDR unlocks the FluxGit-powered tier. Without it, the free-shell tier still works.

FLUXGIT_MCP_AUDIT_LOG enables an append-only JSONL audit log of every tools/call. Arguments are hashed; raw paths and identifiers are never written verbatim.

Semantic diff contract

diff.semantic is the most-used tool for AI agents and the easiest to misuse. The rule is strict:

A result may only be called semantic if data.supported is exactly true.

When the FluxGit semantic engine is not available, diff.semantic returns:

{
  "tool": "diff.semantic",
  "readOnly": true,
  "data": {
    "supported": false,
    "fallback": "diff.text",
    "reason": "Semantic diff is not available in local sidecar fallback mode.",
    "textDiffArguments": { "repoPath": "...", "base": "...", "head": "...", "path": "..." }
  }
}

Connected agents must:

1. Call diff.semantic.

2. Read data.supported.

3. If true, use the semantic payload and label results as semantic.

4. If false, call diff.text with data.textDiffArguments and present results as a text-diff fallback.

5. Never infer function- or class-level moves from a text patch alone.

Allowed wording: "FluxGit reported a text-diff fallback for this file". Prohibited wording: "This is a semantic diff" when supported=false.

Audit log

Every tools/call is optionally appended to a JSONL file pointed at by FLUXGIT_MCP_AUDIT_LOG:

{
  "ts": "...",
  "tool": "repo.status",
  "tier": "free" | "fluxgit" | "fluxgit-write-handshake",
  "ok": true,
  "argumentsHash": "sha256:...",
  "repoScope": "...",
  "summary": "...",
  "signature": "base64url-ed25519",
  "signatureKeyId": "1a2b3c4d5e6f7a8b"
}

Sensitive paths and identifiers are hashed, never stored verbatim.

Per-entry Ed25519 signatures (shipped 2026-05-28)

Audit signing is opt-in. When FLUXGIT_MCP_AUDIT_SIGN_KEY points to a PEM PKCS8 Ed25519 private key, every appended entry is signed with that key. Two extra top-level fields are added:

• signature — base64url (no padding) Ed25519 signature over the canonical JSON of the entry without the signature field.

• signatureKeyId — 16-char hex prefix of the matching public key, so rotated keys can co-exist in the same JSONL.

Canonical JSON rule (verifier must match exactly): recursively sort every object's keys lexicographically by UTF-8 byte order; arrays preserve order; strip the signature and signatureKeyId fields; serialize with serde_json's default compact form (no whitespace, no newlines); sign / verify those bytes.

If the env var is unset, audit entries are written in the legacy unsigned format. If the env var points to a missing or invalid key, the sidecar logs a warning to stderr and falls back to unsigned audit — auditing never refuses to record events.

Verifying an audit log

The sidecar binary doubles as a verifier:

fluxgit-mcp-sidecar verify-audit /path/to/mcp.jsonl --pubkey /path/to/install.pub.pem

Output reports the number of verified, failed, unsigned, and malformed entries, plus the 1-indexed line numbers of any failures. Exit code is 0 when every signed entry verifies, 3 when at least one entry failed verification or was malformed, 2 on usage error. Unsigned entries are counted separately and do not fail the run.

Programmatic verification uses the public verify_audit_event_signature(&event_value, &public_key) function on the sidecar crate, so audit-proof tooling can be embedded anywhere. A MissingSignature error means the entry is unsigned (caller's choice how to treat it); Ok(false) means the signature is present but does not verify under the supplied key.

Protocol details

• MCP protocol version: 2024-11-05

• Transport: stdin/stdout (newline-delimited JSON-RPC 2.0). Legacy Content-Length framing also supported.

• Capabilities: tools (listChanged: false).

• 29 tools in tools/list. Read-only tools advertised with annotations.readOnlyHint: true. Write-handshake tools advertised with annotations.readOnlyHint: false.

Error codes:

Status

This is a working MCP server. The read-only surface is implemented and tested. The write-with-UI-handshake protocol is live as of 2026-05-28: all 5 operation.preview.* tools dispatch through the gateway handshake server, render an "🤖 Requested by AI agent" approval card in the FluxGit app, and complete through the existing safety pipeline (restore points + audit). Clients see structured terminal results (completed | rejected | failed | expired) instead of a placeholder error. The contract is forward-stable.

Roadmap

• End-to-end demo video — public recording of the agent-proposes → user-approves → FluxGit-executes loop, captured from a live install. See product/mcp/DEMO_SCRIPT.md for the script.

• Audit log exportable CSV/JSON — shipped: per-entry Ed25519 signing (2026-05-28). Remaining: exportable CSV/JSON and retention policy for the FluxGit app's audit panel.

• HTTP / SSE transport — for cloud / shared MCP host deployments.

• MCP registry entry — submission to the official MCP server registry once the public release ships.

License

Apache-2.0. See LICENSE.

Related

• FluxGit — the desktop app that produces the FluxGit-powered context.

• MCP for agents — feature page — capabilities, write-handshake contract and roadmap.