repoward

A GitHub MCP server that stops AI agents acting on what an attacker's issue told them to do.

Status: alpha. It has not had an independent security review. Read the threat model, especially the part about where it runs.

The problem

An agent with GitHub access can read a malicious public issue, read a private repository, and post the contents in a public pull request. Each call looks allowed on its own, so per-tool permissions do not stop the sequence. This is the attack Invariant Labs published against agents using GitHub's MCP server, and repoward's test suite replays it against real repositories.

Related MCP server: ContextBridge

Flow guard

Before every write, repoward checks what this agent has already read, using facts from GitHub:

This does not depend on the model noticing an attack or on scanning text, so rewording or encoding the data does not get around it. The state survives restarts, and no tool can clear it. A person can:

uvx repoward flow status
uvx repoward flow reset

What it does not do:

• It only sees this server's tools. A shell, a browser or another MCP server is still a way out, so it protects agents whose GitHub access and outbound routes are limited to this server.

• Data can stay in an agent's context after the window (60 minutes by default) ends. Restart a long-running agent to clear its context.

• Reading one private repository and writing to another private repository is allowed, unless you set single_repo: true.

• Repository files and code search results do not count as untrusted input, because they were merged by someone with write access.

• author_association is a coarse signal, and a compromised trusted account counts as trusted.

• Maintainers of public repositories will see more approval requests. That is the cost of the default, and trusted_associations is configurable.

Features

• Policy: default-deny YAML with roles, repository allowlists and path denylists. See policy.example.yaml.

• Single guarded path: every tool attempt, including unknown tools and invalid arguments, passes through one pipeline and is logged.

• Audit log: a hash-chained record of every decision, with optional HMAC and checkpoints.

• Rate limits: per principal and per tool.

• Read tools: github_list_issues, github_get_issue, github_search_code, github_get_file, ci_get_status and ci_get_job_logs.

• Write tools: github_create_branch, github_commit_files, github_create_pull_request and github_comment. Writes only reach branches with the configured prefix (agent/ by default), pull requests are draft only by default, commit sizes are limited, nothing is forced or retried, and text holding a recognizable credential is refused.

• Dangerous tools: github_merge_pull_request and github_delete_branch need human approval. An approval covers one exact call at one exact commit, works once and expires. Merging is off until a policy turns it on.

• GitHub App tokens (optional): each call gets a short-lived token limited to one repository and only the permissions that tool needs, and GitHub enforces those limits itself.

• Pull request review as approval (optional): a merge can be approved by a real GitHub review of the exact commit.

• Pull request receipts (optional): a commit status summarizing how the agent's run has gone.

• HTTP mode (optional): serve several users from one server, each logging in with a JWT from your identity provider.

• Policy replay: see which past decisions a policy edit would change before you deploy it.

• Untrusted content handling: text from GitHub users comes back wrapped in a tag with a random id per response, after invisible characters, hidden HTML comments and images are removed and secrets are redacted. This is a mitigation, not a guarantee: a model can still be fooled, which is why the rules above are enforced outside the model.

How it works

Every tool call takes the same path, including calls to unknown tools and calls with invalid arguments, so each attempt is logged. Tool code is only ever run from this one place.

MCP client
    |  stdio, or HTTP with a JWT checked on every request
    v
repoward
    1. Find the tool and validate its arguments    fail: refused and logged
    2. Check the policy                            default deny
    3. Check the flow guard                        for writes
    4. Check the rate limit
    5. Check for human approval                    if policy or the flow guard asks for it
    6. Log the decision                            if this fails, a write does not happen
    7. Run the tool
    8. Clean, redact, limit and label the output
    9. Log the outcome
    |
    v
GitHub API

Who it is for

Agents whose only route to GitHub is this server, such as CI bots and hosted agents. What repoward can enforce depends on where it runs relative to the agent:

With a shell on the same OS user, repoward only helps an agent that is misled and not trying to get around it. GitHub-side controls, such as a ruleset protecting your default branch, are the backstop in every case.

By default it runs over stdio as one identity. To serve several people from one server, see Serving over HTTP.

Install

Requires uv. Create a starter policy, which also prints a client config to paste:

uvx repoward init --repo your-org/your-repo

The starter policy gives the server read-only tools. Add --role contributor for write tools. Files live in ~/.repoward, or wherever REPOWARD_HOME points. An existing policy is never overwritten.

Run this first, because the server will not start without a policy. Then add it to your MCP client, and use a token set up as described in GitHub token. Avoid committing a config that holds a real token.

claude mcp add --env GITHUB_TOKEN=<your-token> --transport stdio repoward -- uvx repoward serve

In claude_desktop_config.json, under Settings > Developer > Edit Config:

{
  "mcpServers": {
    "repoward": {
      "command": "uvx",
      "args": ["repoward", "serve"],
      "env": { "GITHUB_TOKEN": "<your-token>" }
    }
  }
}

In ~/.cursor/mcp.json, reading the token from your environment:

{
  "mcpServers": {
    "repoward": {
      "command": "uvx",
      "args": ["repoward", "serve"],
      "env": { "GITHUB_TOKEN": "${env:GITHUB_TOKEN}" }
    }
  }
}

In .vscode/mcp.json or your user profile's mcp.json, which prompts for the token once:

{
  "inputs": [
    { "type": "promptString", "id": "github-token", "description": "GitHub token", "password": true }
  ],
  "servers": {
    "repoward": {
      "type": "stdio",
      "command": "uvx",
      "args": ["repoward", "serve"],
      "env": { "GITHUB_TOKEN": "${input:github-token}" }
    }
  }
}

In ~/.codex/config.toml, forwarding the token from your environment:

[mcp_servers.repoward]
command = "uvx"
args = ["repoward", "serve"]
env_vars = ["GITHUB_TOKEN"]

In ~/.gemini/settings.json, reading the token from your environment:

{
  "mcpServers": {
    "repoward": {
      "command": "uvx",
      "args": ["repoward", "serve"],
      "env": { "GITHUB_TOKEN": "$GITHUB_TOKEN" }
    }
  }
}

In opencode.json in your project, or ~/.config/opencode/opencode.json, reading the token from your environment:

{
  "mcp": {
    "repoward": {
      "type": "local",
      "command": ["uvx", "repoward", "serve"],
      "environment": { "GITHUB_TOKEN": "{env:GITHUB_TOKEN}" }
    }
  }
}

Any other client that can launch a stdio MCP server runs the same command: uvx repoward serve.

If your client reports that the server failed to connect, run uvx repoward serve in a terminal to see why. The usual cause is that repoward init has not been run yet, so there is no policy.

Useful commands:

uvx repoward validate-policy
uvx repoward audit verify

Approving risky actions

A call that needs approval does not run. It returns an approval_id, and a person decides from a terminal:

uvx repoward approvals list
uvx repoward approvals show <id>
uvx repoward approvals approve <id>

The agent then repeats the same call. Pushing new commits changes the commit the call names, so the old approval no longer applies, and GitHub refuses a merge whose head has moved.

Approving with a pull request review

A merge can also be approved by a normal GitHub review, so the approver needs their own GitHub identity and not just access to this machine. Turn it on in your policy:

approvals:
  github_review: { enabled: true, min_permission: write }

A review counts only if it approves the exact commit being merged, comes from a user other than the pull request's author and this server's own identity, and that user has the required access. Any reviewer's outstanding request for changes blocks it. Each check is made again at merge time.

This needs the server to act as a GitHub App or a separate bot account. With your own token, your own reviews are refused. If an agent can reach the reviewer's GitHub credentials, such as a logged-in gh on the same machine, this protection is lost.

Limits of approvals

If the agent runs as the same OS user as this server and has a shell, it could run these commands or edit the database itself. Run the server and its files under a separate user or in a container for approvals to mean anything.

GitHub token

Use a fine-grained personal access token limited to the repositories you allow, with short expiry. Choose Only select repositories so that repository permissions appear. The read tools need these, all read-only:

To use the write tools, also give it read and write access to Contents, Issues and Pull requests. Leave the Workflows permission off: GitHub then refuses workflow file changes even if your policy allows the path.

Contents: write lets a token push to any branch and merge, and GitHub cannot limit it to agent/ branches. That limit lives in this server, so also protect your default branch with a ruleset that requires pull requests and has no bypass for the token's owner. If the token leaks or the server is misconfigured, that ruleset still holds.

Limits to know about:

• ci_get_status does not show third-party checks reported only through the Checks API, which GitHub documents as unavailable to fine-grained tokens.

• github_search_code covers the default branch and files under 384 KB. New repositories can return nothing until GitHub indexes them, and its handling of real results has so far only been tested against simulated responses.

• github_get_file refuses symlinks, submodules, binary files and files over 1 MB. It walks the Git tree, so each read makes one request per folder in the path.

• ci_get_job_logs fails closed if GitHub moves log storage to a host it does not recognize.

• github_merge_pull_request does not support merge queues or stacked pull requests.

• github_delete_branch checks the commit and then deletes, because GitHub has no conditional delete, so a small window remains between the two.

Using a GitHub App instead

A GitHub App gives each tool call its own short-lived token, limited to one repository and that tool's permissions. The App also has its own identity, which review approvals need.

Register an App with webhooks off and only the permissions you need, then install it on the repositories in your policy, and generate a private key:

Then set this in your policy, and point GITHUB_APP_PRIVATE_KEY_FILE at the key in the server's environment:

github_auth:
  mode: app
  client_id: Iv1.your-client-id
  installation_id: 12345678

The private key is a long-lived secret that can issue new tokens. Keep it away from the agent. A contents: write token can still push to any branch of that repository, so the ruleset advice above still applies.

What the audit log proves

The log is tamper-evident, not tamper-proof.

Set REPOWARD_AUDIT_HMAC_KEY to use HMAC, and store the output of repoward audit checkpoint somewhere the server cannot write. Comment bodies, titles, commit messages and file contents are logged as a length and hash rather than as text. A crash during a write leaves an incomplete last line, and new writes are then refused until an operator repairs the log. Secret redaction is best effort.

Serving over HTTP

repoward serve --http serves several users from one server. It acts as an OAuth resource server: your identity provider issues the tokens, and repoward checks one on every request. There is no way to run HTTP mode without login. Add this to your policy:

http_auth:
  issuer: "https://idp.example.com/"
  audience: "https://mcp.example.com/mcp"
  jwks_uri: "https://idp.example.com/.well-known/jwks.json"
  algorithms: [RS256]
  resource_server_url: "https://mcp.example.com/mcp"
  allowed_hosts: ["mcp.example.com"]

identity_map:
  claim: groups
  mapping:
    mcp-readers: reader
    mcp-devs: contributor

repoward serve --http --host 127.0.0.1 --port 8000

• Roles: each caller needs exactly one role. A token whose groups map to no role, or to more than one, is refused, and roles are never combined. The role comes from the current token on each call, so a changed group applies as soon as a new token is used.

• Per user: flow guard labels, rate limits and approvals all belong to the caller, who is identified by the token's iss and sub.

• Audit log: records each caller and their groups, so group names end up in the log. Requests with a bad token are refused before any of this and only logged to stderr, so anonymous traffic cannot grow the audit log.

• Signing keys: come only from jwks_uri, and only asymmetric algorithms are accepted.

• Network: it listens on 127.0.0.1 by default and refuses other addresses without allowed_hosts. It has no TLS of its own, so put it behind a reverse proxy that handles HTTPS and passes the Host header through.

Limits:

• One GitHub identity: callers' tokens are never sent to GitHub, so all users share the server's GitHub credential, and GitHub's own logs show only that. This server's audit log is the only record of who did what.

• One process: rate limits and legacy sessions live in memory, so run a single instance.

• Approvals: the command line approvals still run on the server's host, so whoever approves needs access there. Review approval works as usual.

• Tested with: locally signed tokens against real HTTP and real GitHub repositories, in both MCP protocol versions. It has not yet been tested with a specific identity provider, so try yours on a test setup first.

Pull request receipts

Reviewers of an agent's pull request can see how its run has gone. Turn this on in your policy:

receipts:
  enabled: true

After each commit or pull request the agent makes, repoward posts a commit status named repoward/receipt:

policy 4c16a8364c1a; audit seq 12-40; 2 denied; 1 escalated; flow labels: yes

That is the policy's hash, the audit records since this server started, how many calls were denied or needed approval, and whether anything restricts the agent's writes. flow labels can also say no, unknown, or off if the flow guard is off.

• It needs Commit statuses: write. A GitHub App uses a separate token holding only that permission for each receipt. A personal token needs the permission on the token itself, so every tool's calls carry it.

• A receipt is audited before it is posted. If posting fails, the agent's commit or pull request still succeeds, and the failure is logged.

• A receipt does not prove a change came through repoward. Anyone with that permission can post the same status, so don't make it a required check.

• Anyone who can see the repository can read it. On a public repository, that includes how many calls were refused and whether flow labels are set, which can hint that the agent read private content.

Trying a policy change

Replay a policy against your audit log to see what it would have done differently:

repoward policy replay --policy new-policy.yaml

3 decisions: 1 unchanged, 1 changed, 1 unknown

Changed:
  seq 2  github_comment  org/app: allow -> approval_required  (flow.untrusted_author)

Unknown (the log does not hold what this decision depends on):
  seq 5  github_commit_files  org/app: allow -> unknown  (it depends on file sizes or text the log does not hold)

It never contacts GitHub and changes nothing, and it refuses a log that fails verification. Add --key-env to name the variable holding your HMAC key if the log uses one.

The log leaves out content, so some answers are not knowable. A decision can depend on a file's exact size, on whether hidden text held a credential, or on flow guard facts that were not recorded, such as the visibility of a repository the log never read. Replay reports those as unknown and does not guess.

Replay assumes the agent would have made the same calls. A real agent that was refused would have done something else, so treat the report as a guide to the policy's effect and not as a prediction.

Contributing and security

See CONTRIBUTING.md. Please report vulnerabilities as described in SECURITY.md.

License

MIT. This is an independent project, not affiliated with or endorsed by GitHub.