Inspect the repository, edit files, run commands, use memory, and call MCP tools. Mutating effects require approval by default.
read · edit · shell · memory · MCPAlpha · local-first · Ollama
local-agent is an alpha terminal coding agent for Ollama. Inspect repositories, edit files, run commands, and call MCP tools in NORMAL; enforce read-only work in PLAN; or send proactive work in AUTO. Start bounded durable goals explicitly with /goal.
Mutating effects require approval by default.
Trust boundaryLocal-first validates endpoints. It is not an operating-system sandbox.Live Ollama inventory
local-agent reads Ollama’s authoritative inventory, groups local and Cloud entries, and routes automatically only among admitted local models.
/api/tagsReadable effects
By default, repository mutations, shell commands, memory changes, and MCP calls require an explicit decision. Durable goals use the same policy.
When approval is required, the card keeps the tool and its arguments together before execution.
go test ./... -racewaiting for decisionStart a goal explicitly with /goal. Its objective, limits, permits, receipts, and recovery remain visible as it advances.
Cortex can provide bounded semantic status. local-agent still owns the approvals, transcript, budgets, and final stop decision.
Inspect the boundary
Ollama and network MCP endpoints must use local-machine addresses under local-only policy.
Built-in file tools canonicalize paths; external reads require a temporary exact-file or directory grant, while writes stay in the startup workspace.
Approved shell commands and trusted STDIO servers remain separate trust decisions.
Focused local tools
Connected integrations, test infrastructure, and related projects keep separate ownership boundaries. MCP-capable does not mean bundled or silently trusted.
Start in a clean worktree
Install the current tagged release, connect Ollama, and begin with a read-only request.
go install github.com/abdul-hamid-achik/local-agent/cmd/local-agent@latest